Cambridge Commentary on EU General-Purpose AI Law

2. Substance

2.1. Article 51(1): Classification of GPAI models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain.

10Article 51(1) introduces two alternative conditions under which a GPAI model shall be classified as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .³⁹ The first condition, under point (a),⁴⁰ requires the model to exhibit high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . The second condition, under point (b),⁴¹ concerns the presence of capabilities or an impact equivalent to those high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , as determined by the Commission on the basis of the criteria set out in Annex XIII.

11In addition to Article 51(1), Article 52 sets out the Commission’s power to designate GPAI models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ,⁴² thereby raising the question of the relationship of Article 52’s designation provisions with the classification conditions under Article 51(1). While this relationship is analysed in-depth elsewhere,⁴³ it is apparent that Article 51(1) was intended to serve as the foundational provision for the classification of a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , as evidenced by the wording of the provision and its positioning at the beginning of the classification rules within Section 1. of Chapter V of the AI Act.⁴⁴ The central importance of Article 51(1) is reflected in the Commission’s Guidelines on the Scope of the Obligations for General-Purpose AI Models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. (“Commission Guidelines”), which state that ‘[f]rom the moment when a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. meets either of the two conditions [under Article 51(1)], the model is classified as a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. and its provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must comply with the relevant obligations’.⁴⁵

2.1.1. Article 51(1)(a)

12Article 51(1)(a) sets out the first alternative condition for the classification of a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , requiring that the model ‘has high impact capabilities evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks’.⁴⁶ Article 3(64) defines high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as ‘capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’.⁴⁷

13The classification of GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. under Article 51(1)(a) is best understood within the overall context of the classification rules under Articles 51 and 52.⁴⁸ Under Article 51(2), a model is presumed to have high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , once the cumulative amount of computation used for its training measured in FLOPs is greater than 10²⁵. Where a GPAI model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. or it becomes known that it will have high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , its provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. has to notify the Commission pursuant to Article 52(1)’s first sentence.⁴⁹ Further, Article 52(2) allows the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to contest the classification of the model when submitting the notification.⁵⁰ Moreover, Article 52(1)’s third sentence allows the Commission to designate a GPAI model of which it has not been notified as ‘presenting systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ’ as a GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .⁵¹ This designation provision is arguably linked to Article 51(1)(a) as well.⁵² While Annex XIII primarily informs the determination of whether a GPAI model may be classified under Article 51(1)(b)’s condition,⁵³ certain criteria in the annex may also inform the assessment of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. under Article 51(1)(a),⁵⁴ particularly those included in points (d) and (e) of Annex XIII.⁵⁵

14A thorough understanding of Article 51(1)(a) requires examining the relationship between high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. and systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . Article 3(65) defines systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. as ‘a risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. that is specific to the high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. of general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. , having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain’.⁵⁶ In this context, Article 51(1)(a) can be read as creating a kind of presumption of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. presence in a GPAI model based on its high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .⁵⁷ This finds support in Recital 112’s first sentence, which states that ‘[a] general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. that meets the applicable threshold for high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. should be presumed to be a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .’ The procedure to contest classification under Article 52’s second and third paragraph would then operate as a mechanism to rebut that presumption.⁵⁸ However, unlike Article 51(2),⁵⁹ Article 51(1)(a) does not expressly characterise the link between high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. and systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. as ‘presumptive’ and Article 52(2) and (3) do not expressly characterise the procedure to contest classification as a ‘rebuttal’ of a presumption.⁶⁰

15The contestation procedure under Article 52(2) and (3) explains why a GPAI model’s classification under Article 51(1)(a) generally does not require an independent determination of whether it comes with risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. that meet Article 3(65)’s definition of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .⁶¹ Rather, where a model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. but exceptionally does not present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. can contest the model’s classification pursuant to Article 52(2).⁶² A similar rationale applies to the assessment of whether a model has a significant impact on the internal market:⁶³ although this criterion is expressly mentioned in Article 3(65)’s definition of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ⁶⁴ as well as in Annex XIII,⁶⁵ Article 51(1)(a)’s wording unambiguously precludes an independent requirement that the model should have a significant impact on the internal market under this provision.⁶⁶ Rather, this assessment may be relevant to the procedure for contesting classification under Article 52(2), as well as designation under Article 51(1)(b).⁶⁷ Similar considerations also apply for other criteria without direct relevance for the model’s capabilities, such as the number of registered end users mentioned under point (g) of Annex XIII.⁶⁸

2.1.1.1. Commission designation in the context of Article 51(1)(a)

16Scholars are divided on whether the classification of GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. under Article 51(1)(a) requires a Commission decision⁶⁹ or whether GPAI models meeting that condition are automatically classified by operation of law.⁷⁰ According to its Guidelines, the Commission does not view designation as a necessary requirement for classification under Article 51(1)(a), setting out that a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must comply with the obligations for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. from ‘the moment when the model meets the condition laid down in Article 51(1), point (a), AI Act’.⁷¹

17An interpretation of Article 51(1)(a) as providing for the automatic classification, by operation of law, of GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. is compelling, as the legislative text – despite some ambiguity in the wording of Article 51(1) – more strongly supports it. Some scholars have argued – based on the German language version of Article 51(1) – that the wording ‘shall be classified’ instead of ‘is classified’ or ‘is considered’ implies a further procedural requirement for classification beyond Article 51(1)(a)’s substantive requirement of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .⁷² However, this interpretation is largely undermined by Article 51(1)(b), which expressly requires that determination of equivalent capabilities or impact must be ‘based on a decision of the Commission’. The express provision for a Commission decision in Article 51(1)(b)’s classification condition indicates that no such requirement exists for Article 51(1)(a).⁷³ This interpretation is reinforced by Recital 111’s second sentence, which states that ‘a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. should be considered to present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. if it has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ’ without mentioning the requirement of a Commission decision in this context.

18A comparison with the Digital Markets Act (“DMA”) and the Digital Services Act (“DSA”) reinforces this reading.⁷⁴ Article 51(1)’s use of ‘shall be classified’ notably deviates both from the use of ‘shall be designated’ in the DMA’s provisions for gatekeeper designation and the use of ‘shall […] adopt a decision designating as […]’ in the DSA’s provisions for designation of very large online platforms (“VLOPs”) and very large online search engines (“VLOSEs”).⁷⁵ This terminological difference between Article 51(1) on one hand and Articles 3(1) DMA and 33(4) DSA on the other hand can plausibly be explained by the AI Act’s need to accommodate both for classification by operation of law under Article 51(1)(a) and classification by Commission decision under Article 51(1)(b).

19Beyond arguments grounded in Article 51(1)’s wording, the apparent lack of a legal basis for a designation decision where the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. has notified the Commission of a model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. further supports automatic classification under Article 51(1)(a). While Article 52 does contain two designation provisions, neither of those allow the Commission to designate a GPAI model of which it has been notified solely on the basis of its high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . In fact, Article 52(1)’s third sentence permits designation only of GPAI models ‘of which [the Commission] has not been notified’,⁷⁶ while Article 52(4)’s first subparagraph addresses the designation of a GPAI based on the criteria set out in Annex XIII – and not on the model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .⁷⁷ This, too, strongly suggests that Article 51(1)(a) does not require a Commission decision for classification of a GPAI model with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as a GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .

20The counterarguments against automatic classification under Article 51(1)(a) are not entirely convincing. Some authors contend that the vagueness of Article 51(1)’s substantive requirements and the complexity of their assessment causes legal uncertainty for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. that would not permit automatic classification.⁷⁸ This concern appears reasonable with respect to classification under Article 51(1)(b) which generally requires ‘an overall assessment’ of the criteria set out in Annex XIII⁷⁹ – and can indeed prove challenging to perform for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. . However, the argument is far less strong for classification under Article 51(1)(a) based on a model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . For this classification condition, Article 51(2) provides a presumption of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. based on the cumulative amount of computation used for training. Given the existing methods for estimating training compute⁸⁰ and the Commission Guidelines’ recognition that some leeway is appropriate in making estimates, to ‘account for the difficulties providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may face’,⁸¹ a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. can determine without excessive difficulty whether its model meets the training compute threshold.⁸² Moreover, other provisions, such as the notification obligation under Article 52(1)’s first and second sentence, demonstrate that the AI Act expects providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to self-assess whether their models have high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .⁸³ Article 51(1)(a) and (3) further envisage the Commission’s adoption of indicators and benchmarks to help evaluate the model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. via delegated act,⁸⁴ which may further facilitate such assessments in the future.

21A second argument against automatic classification under Article 51(1)(a) posits that the possibility to contest classification under Article 52(2) can be plausibly explained only if classification under Article 51(1)(a) requires a Commission designation.⁸⁵ This argument relies on the fact that the AI Act does not expressly provide an exemption from classification if the Commission accepts a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s arguments submitted pursuant to Article 52(2) to contest classification.⁸⁶ While Article 52(3) expressly addresses only the rejection, and not the acceptance, of a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s arguments and its consequences, this omission, however, does not support the conclusion that classification under Article 51(1)(a) requires designation. The necessity of an acceptance decision is independent of any requirement for a designation decision for classification under Article 51(1)(a).⁸⁷ Moreover, Article 52(3) does not empower the Commission to designate a GPAI model where the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s contestation of classification fails to meet the requisite standard. Rather, in such cases the Commission ‘reject[s]’ the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s arguments and the GPAI model ‘shall be considered’ to be a GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .⁸⁸ Furthermore, there appears to be no basis to infer a requirement for Commission designation under Article 51(1)(a)’s classification condition from the absence of provisions governing the acceptance of a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s contestation pursuant to Article 52(2), particularly given that – as demonstrated above – such a decision lacks an express legal basis.

22Even though, proceeding from the above, classification under Article 51(1)(a) does not require a Commission designation decision, the Commission can nevertheless designate a GPAI model with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as a GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. under Article 52(1)’s third sentence when it has not been notified of it.⁸⁹ Although this provision refers to a ‘ general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. presenting systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ’ (emphasis added), there are convincing arguments to read designation under Article 52(1)’s third sentence in the context of the rest of Article 52(1), which also relates to GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. – as laid out in further detail elsewhere.⁹⁰

2.1.1.2. High-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models.

23While discussed in-depth elsewhere,⁹¹ Article 3(64) is a key provision to thoroughly understand the concept of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . It defines them as ‘capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’. The omission of the hyphen in Article 51(1)(a) and (2)’s spelling of ‘high impact capabilities’ appears to be an unintended drafting inconsistency without substantive relevance that does not prevent the definition of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. under Article 3(64) from applying in the context of Article 51(1)(a).⁹²

24Article 3(64)’s definition raises several interconnected interpretive questions that are touched upon throughout this chapter. For example, what constitutes the most advanced GPAI models – is this determined domain-specifically (such as most advanced in coding or video generation) or by overall advancement,⁹³ and if the latter, how would overall advancement be assessed? Once the most advanced GPAI models are identified, which capabilities should be compared – must a model match or exceed all capabilities or only relevant ones, and which capabilities would be relevant?⁹⁴ Does Article 3(64)’s reference to ‘recorded’ capabilities further limit the capabilities under consideration?⁹⁵

2.1.1.3. Appropriate technical tools and methodologies, including indicators and benchmarks

25Whether a GPAI model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. must be evaluated on the basis of ‘appropriate technical tools and methodologies, including indicators and benchmarks’.⁹⁶ This enumeration must be interpreted in light of the AI Act’s reference to a multitude of such assessment instruments in the context of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. classification, such as ‘(technical) tools’,⁹⁷ ‘methodologies’,⁹⁸ ‘indicators’,⁹⁹ ‘benchmarks’,¹⁰⁰ ‘criteria’,¹⁰¹ ‘thresholds’,¹⁰² ‘approximations’¹⁰³ and ‘evaluations’,¹⁰⁴ without defining them or clearly distinguishing between them – and which it may be using synonymously.¹⁰⁵ The precise distinction between the assessment instruments mentioned in Article 51(1)(a), as well as their differentiation from the other assessment instruments, is not entirely clear. This lack of clear delineation appears ultimately immaterial for Article 51(1)(a), however, as the terms ‘technical tools’ and ‘methodologies’ appear sufficiently broad to encompass, in principle, all conceivable types of assessment instruments for high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .¹⁰⁶ Moreover, Article 51(1)(a)’s enumeration of indicators and benchmarks as examples of technical tools and methodologies is non-exhaustive (‘including’), such that thresholds¹⁰⁷ and evaluations for capabilities, for instance, can be subsumed under Article 51(1)(a)’s methodologies.

26More practically significant than the precise distinction between different assessment instruments is the question of what requirements must be satisfied for such instruments to qualify as ‘appropriate’ within the meaning of Article 51(1)(a). Recital 111 provides an initial indication as it lays out that ‘[t]hresholds, as well as tools and benchmarks for the assessment of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , should be strong predictors of generality, its capabilities and associated systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. of general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. , and could take into account the way the model will be placed on the market Article 3(9) AI Act: ‘placing on the market’ means the first making available of an AI system or a general-purpose AI model on the Union market. or the number of users it may affect’ (emphasis added).¹⁰⁸ The critical consideration appears to be whether the respective assessment instruments are sufficient predictors of the model’s capabilities – and not the model’s generality or its systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. also mentioned by the recital – as it is those capabilities that trigger classification under Article 51(1)(a).

27Interestingly, the Safety and Security Chapter of the Code of Practice regards ‘appropriate’ as a less demanding standard compared to ‘best practice’¹⁰⁹ or ‘state of the art’,¹¹⁰ defining it as ‘suitable and necessary to achieve the intended purpose Article 3(12) AI Act: ‘intended purpose’ means the use for which an AI system is intended by the provider, including the specific context and conditions of use, as specified in the information supplied by the provider in the instructions for use, promotional or sales materials and statements, as well as in the technical documentation. of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. assessment and/or mitigation, whether through best practices, the state of the art, or other more innovative processes, measures, methodologies, methods, or techniques that go beyond the state of the art’ (emphasis added).¹¹¹ The definitions contained in the Code of Practice are not directly applicable to the AI Act but they can contribute to its literal and systematic interpretation.¹¹²

28Further drawing on the Safety and Security Chapter of the Code of Practice, relevant criteria for assessing a method’s appropriateness could include its scientific rigour, including its validity and reproducibility.¹¹³ Arguably, assessment instruments that have comparable or superior predictive power to training compute thresholds should be permissible, as the legislature has deemed such thresholds a sufficient basis for Article 51(2)’s presumption of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. and thus appropriate. Article 51(3) empowers the Commission to introduce and specify the instruments relevant for assessing a model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .¹¹⁴

29The AI Act does not, however, require the formal adoption of technical tools and methodologies for evaluating a model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. via a delegated act.¹¹⁵ Such a requirement does not find support in Article 51(1)(a)’s wording, which requires only that the assessment instruments be ‘appropriate’ but not that they be provided for in a delegated act. In this respect, Article 51(1)(a) differs from other provisions where the legislature expressly set out that their applicability is contingent upon the adoption of a delegated act.¹¹⁶ Nor does such a requirement follow from Article 51(3). Even if this provision, by its wording (‘shall’), establishes the Commission’s obligation to adopt delegated acts in certain circumstances,¹¹⁷ this does not support the requirement that assessment instruments must be formally adopted under Article 51(1)(a). This is because the obligation under Article 51(3) is limited to instances where the adoption of a delegated act is necessary for the thresholds under Article 51(1) and (2) to reflect the state of the art.¹¹⁸

2.1.2. Article 51(1)(b)

30Article 51(1)(b) contains the second alternative condition for the classification of a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . This condition comprises two main requirements. First, it requires that the GPAI model has ‘capabilities or an impact equivalent to those set out in point (a)’ of Article 51, that is, high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .¹¹⁹ Annex XIII contains criteria that should be taken into account in this context.¹²⁰ Second, this determination must be ‘based on a decision of the Commission, ex officio or following a qualified alert from the scientific panel’.¹²¹ This requirement refers to the need for a designation decision pursuant to Article 52(4)’s first subparagraph for classification under Article 51(1)(b).¹²²

31The classification condition under Article 51(1)(b) serves ‘to complement’ the system of high-impact capabilities-based classification under Article 51(1)(a) by allowing the Commission ‘to take individual decisions designating a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. as a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ’ on the basis of an ‘overall assessment’ of Annex XIII criteria.¹²³ This complementary function may operate in several ways. First, whereas Article 51(1)(a) focuses on a model’s capabilities,¹²⁴ Article 51(1)(b) allows for classification based on a model’s impact.¹²⁵ The provision’s reference to a model’s impact alongside its capabilities appears to reflect the legislature’s conception that ‘a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. should be considered to present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. if it has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. […] or significant impact on the internal market due to its reach.’ (emphasis added).¹²⁶ Second, to the extent that classification under Article 51(1)(a) may not account for models with particularly high capabilities in a specific domain such as offensive cyber capabilities without high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. across all domains, Article 51(1)(b) could fill these gaps.¹²⁷ It has further been proposed that Article 51(1)(b) could allow the classification of a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. where its provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. technically reduces the amount of computation used for the model’s training to fall below Article 51(2)’s threshold without this being reflected in the model’s actual capabilities.¹²⁸

2.1.2.1. Capabilities or an impact equivalent to those set out in point (a)

32Article 51(1)(b) allows for classification of a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. where it has ‘capabilities or an impact equivalent to those set out in point (a)’, meaning that classification under Article 51(1)(b) requires the model to have capabilities or an impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .¹²⁹ In this context, some legal scholars have argued that both Article 51(1)(a) and Article 51(1)(b) impose equal substantive requirements, with the implication that Article 51(1)(b) – despite its different wording – requires the model to possess high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , just as Article 51(1)(a) does.¹³⁰ On this view, the difference between point (a) and (b) of Article 51(1) is primarily procedural in nature and would not establish substantively different thresholds for systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. classification under the two alternatives.¹³¹ As will be examined below, this does not appear entirely convincing.¹³² Rather than mirroring Article 51(1)(a)’s substantive requirements, Article 51(1)(b) establishes a condition for classification in its own right. The interpretation of this condition raises several questions: what is the role of Annex XIII for the assessment under Article 51(1)(b)?¹³³ When are a model’s capabilities equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ?¹³⁴ What constitutes an impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ?¹³⁵ Does the cumulative equivalence of a model’s capabilities and impact to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. suffice?¹³⁶ How does the proven presence or absence of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. influence classification under Article 51(1)(b)?¹³⁷ These questions are examined in the sections that follow.

33Legal scholars have rightly noted that the criteria for classification under Article 51(1)(b) and Annex XIII leave considerable room for flexibility, such that the Commission may enjoy substantial discretion in designating a GPAI model on this basis.¹³⁸ The existence of a margin of discretion finds support in the wording of the designation provision under Article 52(4)’s first subparagraph (‘may’) which relates to Article 51(1)(b).¹³⁹ Moreover, Article 51(1)(b)’s wording itself suggests a substantial margin of discretion for the Commission for deciding whether the requirements for designation are met.¹⁴⁰ Particularly striking is the fact that the requirement of a Commission decision is placed before the substantive requirements in the provision’s structure (‘based on a decision of the Commission, […] it has capabilities or an impact equivalent to those set out in point (a)’),¹⁴¹ thereby creating the impression that the focus lies on the Commission’s assessment rather than the objective presence of capabilities or an impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . It may be this wording which has led some legal scholars to characterise Article 51(1)(b) as establishing ‘subjective classification’.¹⁴² Furthermore, Recital 111 refers to the Commission’s ‘overall assessment’ of Annex XIII criteria in the context of Article 51(1)(b), without prescribing any weighting of these criteria.¹⁴³ In light of Article 51(1)(b)’s complementary purpose¹⁴⁴ and the legislature’s concerns regarding technological developments,¹⁴⁵ it appears close at hand that the legislature, by creating a more flexible Article 51(1)(b) with broad Commission discretion alongside Article 51(1)(a), sought to ensure that the classification framework provided by both provisions would be future-proof and resilient to disruption.¹⁴⁶

34Before turning to Article 51(1)(b)’s precise requirements,¹⁴⁷ it is worth considering whether the substantive requirements for classification under Article 51(1)(b) are distinct from those under Article 51(1)(a).¹⁴⁸ Several considerations support the view that Article 51(1)(b) establishes distinct substantive requirements. First, the provisions employ different language: while Article 51(1)(a) refers to ‘ high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ’, Article 51(1)(b) requires ‘capabilities or an impact equivalent to those set out in point (a)’.¹⁴⁹ Second, classification under Article 51(1)(a) is directly based only on the model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , including indicators and benchmarks for such capabilities, whereas Article 51(1)(b) requires the Commission to have regard to the criteria in Annex XIII,¹⁵⁰ some of which concern the model’s reach¹⁵¹ – such as the number of registered business and end users (points (f) and (g)) – rather than its capabilities. These criteria serve as indicators for the model’s impact, and do not indicate, or indicate only tangentially, whether a model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .¹⁵²

35Third, training compute plays a fundamentally different role under each condition. Under Article 51(1)(a), exceeding the training compute threshold in Article 51(2) alone creates a presumption of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ¹⁵³ and is therefore sufficient for classification.¹⁵⁴ Under Article 51(1)(b), however, the amount of training compute is merely one criterion among several listed in Annex XIII that inform the model’s classification,¹⁵⁵ an approach which likely aims to enable classification of models staying below the Article 51(2) threshold.¹⁵⁶ Fourth, Article 52 contains two distinct provisions for designation of GPAI models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ,¹⁵⁷ one arguably permitting designation on the basis of Article 51(1)(a) and the other on the basis of Article 51(1)(b).¹⁵⁸ If both conditions applied the same substantive standard, the existence of two separate designation procedures would be difficult to explain. Finally, if Article 51(1)(b) merely replicated Article 51(1)(a)’s substantive requirements, it would risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. being redundant:¹⁵⁹ models meeting those requirements would already automatically be classified under Article 51(1)(a).¹⁶⁰

36An interpretation of Article 51(1)(b) as permitting classification without requiring high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. can be reconciled with Article 3(65)’s definition of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .¹⁶¹ If systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. refers to risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. that is ‘specific to the high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. of general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’ (emphasis added), the question arises whether the classification of GPAI models without high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. can be justified. Such classification is possible. As discussed below, it is uncertain whether Article 3(65) necessarily implies that systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. can only arise in GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. or whether it is merely characteristic of models with such capabilities.¹⁶² But even in the former case, the proven absence of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. is not the same as uncertainty about their presence. Accordingly, the fact that Article 51(1)(b) does not require the Commission to establish the model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. for classification does not necessarily mean it permits designation of models proven not to possess such capabilities.¹⁶³ Rather, it permits the Commission to designate a GPAI model according to a distinct substantive standard – capabilities or an impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. – even in cases where evidence concerning the presence of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. is inconclusive. After a minimum period of six months, a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. can request reassessment of its model’s designation on the basis of new reasons that have arisen since the designation decision.¹⁶⁴

37As indicated above, some legal scholars nevertheless contend that the conditions in point (a) and (b) of Article 51(1) do not entail different substantive requirements for classification.¹⁶⁵ They argue that Article 51(1)(b)’s express reference to Article 51(1)(a) indicates substantive equivalence, and that the difference between these conditions can be explained on procedural grounds.¹⁶⁶ According to this view, the distinction between Article 51(1)(a)’s evaluation of the model’s capabilities on the basis of appropriate technical tools and methodologies and Article 51(1)(b)’s determination with regard to the criteria set out in Annex XIII merely reflects the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s duty to conduct an actual evaluation of its model.¹⁶⁷ This procedural explanation is not entirely convincing,¹⁶⁸ as procedural reasons may explain Article 51(1)(a)’s reference of ‘appropriate technical tools and methodologies’ and Article 51(1)(b)’s reference of Annex XIII but cannot explain why Article 51(1)(b) requires ‘capabilities or an impact equivalent to those set out in point (a)’ rather than simply ‘ high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ’.

38Annex XIII contains criteria¹⁶⁹ which are particularly relevant for determining whether a model has capabilities or an impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .¹⁷⁰ The importance of these criteria for this assessment under Article 51(1)(b) is evidenced by the provision’s express reference to Annex XIII and reinforced by the introductory sentence of this annex, which expressly states that it contains a list of criteria that the Commission shall take into account ‘[f]or the purpose of determining that a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. has capabilities or an impact equivalent to those set out in Article 51(1), point (a)’.

39The subsequent sections discuss whether and to what extent the Commission is required to take the criteria contained in this annex into account when designating a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. on the basis of Article 51(1)(b) in conjunction with Article 52(4)’s first subparagraph,¹⁷¹ and argue that the list of criteria contained in Annex XIII is non-exhaustive in this context.¹⁷²

40Multiple provisions suggest a duty for the Commission to consider the criteria contained in Annex XIII when designating a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. on the basis of Article 51(1)(b) in conjunction with Article 52(4)’s first subparagraph.¹⁷³ Annex XIII itself establishes that the Commission ‘shall take into account [its] criteria’ for the purpose of determining that a GPAI model has capabilities or an impact equivalent to those set out in Article 51(1)(a).¹⁷⁴ Article 51(1)(b) reinforces this by requiring the Commission to ‘[have] regard to the criteria set out in Annex XIII’. Article 52(4)’s first subparagraph and Recital 111 further confirm that this decision must be taken ‘on the basis of criteria set out in Annex XIII’ and ‘on the basis of an overall assessment of [those] criteria’ (emphasis added), respectively. The Commission’s duty to consider the Annex XIII criteria is reminiscent of the Commission’s general duty of diligent and impartial examination, which requires it to carefully examine the relevant facts of a case¹⁷⁵ and is – at least for some areas of EU law such as competition law and state aid – recognised by the EU Courts.¹⁷⁶

41This raises the question of how this duty relates to the substantive requirements for designation under Article 51(1)(b), in particular whether the Commission must enquire about all of Annex XIII’s criteria to determine that a GPAI model has capabilities or an impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. or whether it can base its designation decision on only some of the criteria.¹⁷⁷ Compelling arguments can be made in support of both interpretations. In favour of the Commission’s discretion to selectively focus on some criteria, one might argue that Article 51(1)(b)’s use of ‘or’ permits the Commission to base its designation decision either on the model’s impact or on the model’s capabilities and therefore to focus solely on the Annex XIII criteria relating to capabilities while ignoring those concerning impact, or vice versa.¹⁷⁸ For example, where the Commission comes to the conclusion that a GPAI model has capabilities equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. based on Annex XIII’s capabilities-related criteria (points (a)–(e)), one may argue it would not also have to enquire about the model’s number of registered business and end users (Annex XIII, points (f) and (g)). This would appear sensible, insofar as the number of registered users appears to be a rather irrelevant approximation for a model’s capabilities in many cases.¹⁷⁹ However, this selective approach sits in tension with the abovementioned provisions signalling a duty to take Annex XIII criteria into account. Given that Annex XIII establishes that the Commission ‘shall’ rather than ‘may’ take the criteria into account and Recital 111 refers to an ‘overall assessment’,¹⁸⁰ this implies mandatory consideration of all Annex XIII criteria.¹⁸¹

42Two arguments nevertheless support a flexible interpretation of the Commission’s duty to consider Annex XIII criteria, such that the Commission enjoys discretion in its decision over which Annex XIII criteria to focus on¹⁸² – just as Article 51(1)(b) allows the Commission to consider either capabilities or impact.¹⁸³ First, the wording of ‘take into account’,¹⁸⁴ ‘having regard to’¹⁸⁵ and ‘on the basis of’¹⁸⁶ does not appear to impose a particularly stringent requirement. In particular, it does not imply that any single criterion necessarily qualifies or disqualifies a model with regard to Article 51(1)(b)’s standard.¹⁸⁷ Rather, it appears sufficient to engage in a reasoned analysis demonstrating that a particular criterion does not, in the individual case, contribute to the decision on whether Article 51(1)(b) is met, thereby having regard to that criterion. Second, an overly strict interpretation of the Commission’s duty to consider Annex XIII criteria risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. creating perverse incentives for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. . Such an interpretation could encourage providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to withhold information – such as the precise number of parameters – when responding to requests from the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. for the model’s technical documentation¹⁸⁸ or other information,¹⁸⁹ thereby preventing designation of their model through non-cooperation. Nevertheless, where information regarding a relevant criterion is lacking, the Commission will generally be obliged to afford the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. an opportunity to supply it, as follows from the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s right to be heard before a designation decision is made.¹⁹⁰

43A further question concerns whether Annex XIII is exhaustive – that is, whether the Commission may also take into account additional criteria beyond those listed, such as the model’s architecture, training methods, or foreseeable negative effects, when designating a model under Article 51(1)(b) in conjunction with Article 52(4)’s first subparagraph.¹⁹¹ The more compelling view is that Annex XIII is non-exhaustive.¹⁹² In the absence of other conclusive textual indicators,¹⁹³ both Article 51(1)(b)’s wording (‘having regard to’) and Annex XIII’s wording (‘take into account’) suggest that other criteria may be taken into account for designation under Article 51(1)(b) in conjunction with Article 52(4), first subparagraph.¹⁹⁴ Moreover, the inclusion of broad and open-ended criteria in Annex XIII – with non-exhaustive lists of examples¹⁹⁵ – suggests that the legislature sought to ensure consideration of diverse factors rather than to confine the Commission to a closed list.¹⁹⁶

44This interpretation finds further support in the fact that relevant criteria are not mentioned in the list. In particular, while Annex XIII expressly mentions the number of parameters, it does not mention model architecture.¹⁹⁷ This does not appear to be a conscious omission, as both the number of parameters and model architecture are listed together in point 1(d) of Section 1. of Annex XI as part of the technical documentation referred to in Article 53(1)(a) which providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must, upon request, provide to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. .¹⁹⁸ Moreover, information about the model’s architecture provides helpful context for information about the number of parameters.¹⁹⁹ Further, Recital 111’s tenth sentence mentions ‘the way the model will be placed on the market Article 3(9) AI Act: ‘placing on the market’ means the first making available of an AI system or a general-purpose AI model on the Union market. ’ as a relevant criterion for systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. classification, which is not expressly reflected in Annex XIII itself. Similarly, while Article 3(65) defines systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. by reference to both (i) the model’s reach and (ii) its ‘actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or society as a whole’ as indicators of the model’s impact, the Annex expressly mentions only the former and not the latter, even though both appear side by side in that definitional provision. While it might be possible and necessary to interpret Annex XIII criteria extensively to encompass considerations not expressly stated,²⁰⁰ these omissions further suggest that Annex XIII is non-exhaustive.

45The question of Annex XIII’s exhaustiveness is connected to the question of the scope of the Commission’s power to amend Annex XIII under Article 52(4)’s second subparagraph, specifically whether the Commission is permitted to add criteria to Annex XIII.²⁰¹ As laid out elsewhere, there are compelling reasons for interpreting this delegation of power broadly as allowing the Commission to add new criteria.²⁰²

46The capabilities of a GPAI model generally refer to its ability to perform various tasks.²⁰³ Points (a)–(e) of Annex XIII contain criteria particularly relevant for assessing model’s capabilities,²⁰⁴ namely: (a) its number of parameters; (b) the quality and size of its (training) data set; (c) the amount of computation used for its training; (d) its input and output modalities; and (e) its performance on capability benchmarks and evaluations.²⁰⁵ Moreover, the Safety and Security Chapter of the Code of Practice lists capabilities which may contribute to a model presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , including offensive cyber capabilities, capabilities to adaptively learn new tasks or capabilities to evade human oversight.²⁰⁶

47The AI Act does not expressly specify when a model’s capabilities are equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . In line with Article 51(1)(b)’s interpretation as distinct from Article 51(1)(a),²⁰⁷ it is not convincing to equate ‘capabilities […] equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ’ under Article 51(1)(b) with ‘high impact capabilities’ under Article 51(1)(a), as Article 51(1)(b) requires ‘equivalent’ and not ‘the same’ capabilities and such an interpretation would render the inclusion of ‘capabilities’ alongside ‘impact’ in Article 51(1)(b) redundant.²⁰⁸ Given that the standards diverge, the key question becomes how they differ. Any specification of ‘capabilities […] equivalent to [ high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ]’ necessarily presupposes a prior determination of what qualifies as a high-impact capability Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .²⁰⁹

48This question matters for two interconnected reasons. First, the notion of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. forms part of the standard of capabilities ‘equivalent to those set out in point (a)’ under Article 51(1)(b), thus directly determining the provision’s scope. Second, high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. do not only form part of Article 51(1)(b)’s standard but they themselves also form the basis for automatic classification under Article 51(1)(a) and arguably constitute grounds for designation under Article 52(1), third sentence.²¹⁰ Understanding when a model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. is therefore essential to determine the scope of classification under these provisions, and by extension how Article 51(1)(b)’s classification of GPAI models with capabilities equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. could meaningfully complement these provisions.²¹¹

49As laid out above, the definition of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. under Article 3(64) raises various interpretive issues:²¹² for example, what constitutes the ‘most advanced’ GPAI models – is advancement determined in a domain-specific assessment (such as in coding or video generation) or by overall advancement, and if the latter, how is overall advancement assessed? Once the most advanced models are identified, which capabilities should be compared – must a model ‘match or exceed’ all capabilities of these models, or only relevant ones, and which capabilities would be relevant? Does Article 3(64)’s reference to ‘recorded’ capabilities further limit the capabilities under consideration? The resolution of these interpretive questions – particularly if narrower interpretations of the notion of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. prevail – could give rise to different ways in which Article 51(1)(b)’s classification of models with capabilities equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. could complement Article 51(1)(a). In principle, a narrow interpretation of Article 51(1)(a) correspondingly expands Article 51(1)(b)’s scope to capture GPAI models that may present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. yet are not classified under Article 51(1)(a).

50Article 51(1)(b) arguably allows for the classification of GPAI models on the basis of domain-specific high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. – understood here as high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. relating to specific environments into which a GPAI model may be deployed – including through integration into an AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. ²¹³ – which are characterised in particular by their relevant modalities and tasks.²¹⁴ This could include classification of models with particularly high chemical, biological, radiological, and nuclear (“CBRN”) or offensive cyber capabilities. Such classification assumes particular significance if Article 51(1)(a) only encompasses cross-domain high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. – a reading which remains uncertain but appears plausible given Article 3(64)’s wording.²¹⁵

51Annex XIII supports classification under Article 51(1)(b) on the basis of domain-specific high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . While Annex XIII could in principle be used to assess both overall high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. and domain-specific high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , two features of this annex indicate that the legislature drafted it with domain-specific assessments of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. in mind. First, point (d) of Annex XIII refers to the ‘specific type of inputs and outputs (e.g. biological sequences)’ (emphasis added) as a relevant criterion distinct from the broader criterion of ‘input and output modalities of the model’. If only cross-domain capabilities mattered, the mentioning of specific types of input and output would be redundant, as all modalities would necessarily be relevant – a point already conveyed by the broader criterion.²¹⁶ This suggests that specific types of input and output and the corresponding domain-specific capabilities to process them matter in themselves. Moreover, the example of ‘biological sequences’ appears specifically tailored to a domain-specific capability assessment, namely that of biological capabilities.²¹⁷ Second, point (d) of Annex XIII refers to ‘state of the art thresholds for determining high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. for each modality’ (emphasis added) as a further relevant criterion. Meeting a modality-specific threshold would carry little significance in itself if only a model’s cross-domain high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. mattered.²¹⁸ In addition, given that point (e) of Annex XIII mentions ‘benchmarks and evaluations of capabilities’ separately from the modality-specific thresholds in point (d), this distinction further indicates that domain-specific capabilities are relevant in themselves rather than serving merely as indicators of a model’s overall capabilities.

52Moreover, domain-specific capabilities such as CBRN or offensive cyber capabilities²¹⁹ appear to be particularly relevant to certain systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. a GPAI model may cause, such as the risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. of biological or cyber-attacks.²²⁰ Were these not covered by either Article 51(1)(a) or Article 51(1)(b), some GPAI models that present such risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. would fall outside Article 51(1)’s classification framework and corresponding systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. would be left unaddressed. However, that the legislature was aware of such domain-specific capabilities and intended them to be covered finds support in point (d) of Annex XIII mention of ‘biological sequences’, an example that fits a domain-specific assessment of biological capabilities.²²¹

53On the basis that Article 51(1)(b) allows for a model’s classification solely based on domain-specific high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , further questions arise as to which domains are relevant and how one determines high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. in these domains. Given Article 51(1)(b)’s integration into the context of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. classification, it appears appropriate to focus on capabilities that are particularly relevant to the systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. that GPAI models may pose.²²² In that regard, the fourteen model capabilities listed in the Safety and Security Chapter of the Code of Practice as potential sources of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. merit particular consideration.²²³ These include not only the aforementioned offensive cyber and biological capabilities but also further capabilities such as the capacity for manipulation, autonomous operation, and the evasion of human oversight.²²⁴ Further regard should be had to how the AI Act addresses domain-specific capabilities. Some capabilities are mentioned expressly, such as ‘offensive cyber capabilities’.²²⁵ Others are addressed indirectly: biological capabilities, for instance, are referenced through Recital 110’s mention of biological risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. and Annex XIII’s reference to biological sequences as model input. Moreover, international approaches that identify certain capabilities as particularly relevant to systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. could also be taken into account.²²⁶

54The determination of whether a GPAI model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. in a specific domain requires appropriate assessment instruments, just as does the determination under Article 51(1)(a).²²⁷ This likely includes the use of domain-specific capabilities benchmarks such as the Biological Laboratory Protocol Benchmark (BioLP-bench)²²⁸ and the Language Agent Biology Benchmark (LAB-Bench)²²⁹ for AI models’ biological capabilities.²³⁰ Article 3(64)’s standard of ‘capabilities recorded in the most advanced general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’ can guide the determination of which domain-specific capabilities are sufficient to satisfy Article 51(1)(b)’s standard of ‘capabilities […] equivalent to [ high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ]’ as the most advanced models can be determined both across domains and domain-specifically.²³¹

55Beyond allowing for the classification of models with domain-specific high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , Article 51(1)(b)’s reference to capabilities equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. could complement classification under Article 51(1)(a) in further ways.²³² While a detailed analysis of Article 3(64)’s definition of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. – which would be necessary to examine all possible ways in which Article 51(1)(b) could complement Article 51(1)(a) – exceeds the scope of this chapter,²³³ one possibility appears worth discussing: whether Article 51(1)(b) establishes a lower capabilities threshold than Article 51(1)(a) does, such that capabilities insufficient for automatic classification could nonetheless suffice for designation standing alone, without additional offsetting factors such as the model’s high impact.²³⁴

56Such an interpretation, while ultimately less convincing, finds some support in the fact that classification under Article 51(1)(b), in contrast to classification under Article 51(1)(a),²³⁵ requires a Commission designation decision,²³⁶ a requirement which arguably compensates for a lower substantive threshold by providing legal certainty with regard to a model’s classification. It would read ‘equivalent’ as a deliberate choice of wording over ‘equal’, thus permitting a potentially different (lower) level of capabilities.

57However, several considerations weigh against this interpretation. While the wording of Article 51(1)(b) does not preclude it, neither does it provide strong support, as ‘equivalent’ does not necessarily mean that lesser capabilities suffice. More importantly, treating capabilities below Article 51(1)(a)’s threshold as sufficient for Article 51(1)(b) appears unjustified given the substantively identical effect, in particular the application of Article 55(1) obligations, that follows from classification under both provisions.²³⁷ Additionally, the reference in point (d) of Annex XIII to modality-specific high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. signals that the threshold of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. retains relevance for Article 51(1)(b), even if only with regard to specific domains. The more compelling view is therefore that capabilities below the Article 51(1)(a) threshold are not sufficient for meeting Article 51(1)(b)’s standard of capabilities equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. in themselves. Rather, such lower capabilities would need to be offset by other factors, including domain-specific high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ,²³⁸ a particularly high impact²³⁹ or the prospect that the model will predictably reach high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. soon.

58Article 51(1)(b) allows for classification based not only on a GPAI model’s “equivalent capabilities”²⁴⁰ but also based on its impact. The impact of a GPAI model can be described in terms of its reach, including the number of its users, and in terms of any actual or reasonably foreseeable negative effects that stem from it.²⁴¹ Points (f) and (g) of Annex XIII contain criteria particularly relevant for assessing model’s impact.²⁴²

59It has been questioned whether a model’s impact alone can form the basis for a model’s classification as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. under Article 51(1)(b),²⁴³ as Article 3(65) defines systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. as ‘specific to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ’ and not as specific to a model’s high impact.²⁴⁴ In particular, it has been argued that a model’s widespread use alone cannot be sufficient for classification.²⁴⁵ While this is ultimately correct, it must be noted, however, that a model’s reach is indeed a relevant factor to determine a model’s risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. , as widespread use statistically increases the overall probability of harm occurring.²⁴⁶ Indeed, according to Recital 110 and the Safety and Security Chapter of the Code of Practice, the risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. associated with a model increases both with its capabilities and with its reach. At the same time, the reference to actual or reasonably foreseeable negative effects alongside reach as contributing factors that determine a model’s impact in Article 3(65) demonstrates that the legislature does not equate impact with reach.²⁴⁷ This suggests that a model lacking both high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. and the ability to produce significant negative effects on the protected interests set out in Article 3(65)²⁴⁸ cannot be classified as a GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. under Article 51(1)(b), even if it is widely used and therefore has a particularly high reach.

60Regarding the standard for impact-based classification under Article 51(1)(b), the AI Act does not expressly specify when a model’s impact is ‘equivalent to those set out in point (a)’. While Article 51(1)(a) sets out the standard of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , a direct comparison of a model’s impact with this capabilities standard is not readily feasible. Therefore, it appears that the legislature used ‘an impact equivalent to those set out in point (a)’ as shorthand for ‘an impact of a GPAI model being equivalent to the impact of GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ’.²⁴⁹

61On this basis, Article 51(1)(b) requires the determination of an ‘impact standard’ of GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. against which models to be classified under Article 51(1)(b) can be evaluated to determine their impact equivalence. On the basis of Article 3(64) and (65)’s definitions of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. and systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ,²⁵⁰ this likely involves a comparison of the model with the most advanced GPAI models in terms of their reach and their actual and reasonably foreseeable negative effects. The question of whether Article 51(1)(b)’s use of ‘equivalent’ requires a model to at least match the impact of GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ,²⁵¹ or whether – since it does not require an ‘equal’ impact – it sets a slightly lower threshold, likely has less practical relevance, given that such distinctions are difficult to draw for a qualitative threshold such as impact.

62The business user threshold in point (f) of Annex XIII is likely to assume considerable importance for impact-based classification under Article 51(1)(b). According to this provision, a model is presumed to have a high impact on the internal market due to its reach ‘when it has been made available to at least 10 000 registered business users established in the Union’.²⁵² A model that reaches this threshold and is therefore presumed to have a high impact generally satisfies the requirements to be classified as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. under Article 51(1)(b).²⁵³ A high impact linguistically forms part of the notion of ‘ high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ’ which suggests – assuming that the legislature chose the term ‘ high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ’ deliberately²⁵⁴ – that a model with a (presumed) high impact also has an impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. under Article 51(1)(b).²⁵⁵ Recital 111 confirms this interpretation, stating that ‘a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. should be considered to present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. if it has […] significant impact on the internal market due to its reach’.²⁵⁶ The fact that this threshold, unlike the training compute threshold under Article 51(2), is not contained in Article 51 itself²⁵⁷ and does not precisely correspond to Article 51(1)(b)’s wording of ‘an impact equivalent to those set out in point (a)’ does not undermine this conclusion, as the criteria in Annex XIII, and thus the high-impact presumption under point (f) of Annex XIII, evidently relate to Article 51(1)(b),²⁵⁸ and the presumption would otherwise appear to lack a meaningful scope of application.²⁵⁹

63The AI Act does not specify whether the presumption of a model’s high impact under point (f) of Annex XIII is rebuttable. Its refutability is supported by the fact that a high number of registered business users does not necessarily indicate an impact sufficient to assume the presence of systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , since reach is, as laid out above,²⁶⁰ only one relevant factor to determine a model’s impact.²⁶¹ Moreover, an irrefutable presumption would conflict with the Commission’s duty to have regard to Annex XIII for designation under Article 51(1)(b) in conjunction with the first subparagraph of Article 52(4).²⁶² This duty likely implies that the Commission must consider all of the Annex XIII criteria even if it does not decisively base its decision on all of them,²⁶³ thereby precluding any single criterion – as would be the case with an irrefutable presumption – from automatically qualifying the model for designation.²⁶⁴

64The business user threshold in point (f) of Annex XIII is not the only criterion the Commission has to take into account for impact-based classification under Article 51(1)(b). Point (g) of Annex XIII mentions an additional reach-related criterion – the model’s number of registered end users – but refrained from combining this criterion with a presumption or any other threshold indicating a relevant number of registered end users.²⁶⁵ Interestingly, Annex XIII does not contain any criterion specifically aimed at determining the actual or reasonably foreseeable negative effects stemming from a model. These effects may require consideration under point (g) of Annex XIII mentioning ‘evaluations of capabilities of the model’.²⁶⁶ More generally, it appears unlikely that Annex XIII restricts the Commission in its consideration of a model’s effects, as Article 3(65) expressly mentions these in its reference to a model’s impact²⁶⁷ and Annex XIII’s list of criteria is arguably non-exhaustive.²⁶⁸ The assessment of a model’s effects could, for example, be informed by so-called AI incident trackers and databases.²⁶⁹

65The preceding discussion has examined the conditions under which a model’s capabilities or impact might independently satisfy Article 51(1)(b)’s standard of equivalence to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . The question thus arises of whether capabilities and impact, when individually insufficient, could together constitute a basis for equivalence and thereby support classification under Article 51(1)(b) (“cumulative equivalence”).²⁷⁰

66Article 51(1)(b)’s wording is ambiguous in this respect. The phrase ‘capabilities or an impact equivalent to those set out in point (a)’ may be read as referring to either capabilities that are equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. or an impact that is equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . At the same time, the provision’s wording does not exclude a reading under which ‘equivalent to those set out in point (a)’ attaches to ‘capabilities or an impact’ as a composite expression.²⁷¹ On one reading, therefore, at least one element – either capabilities or impact – must independently establish equivalence to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , such that their mere combination cannot suffice where each falls short on its own (“independent equivalence”). On an alternative reading, capabilities and impact may be considered together, with their cumulative effect potentially establishing the requisite equivalence even where neither alone would meet the threshold.²⁷²

67A reading of Article 51(1)(b)’s wording as requiring “independent equivalence” appears more natural. However, the fact that Annex XIII does not expressly distinguish between (i) criteria relevant to assessing whether a GPAI model has sufficient capabilities and (ii) those relevant to assessing whether it has sufficient impact rather supports an interpretation of Article 51(1)(b) of allowing classification in cases of cumulative equivalence. Article 51(1)(b)’s general reference to Annex XIII, combined with the existence of a single undivided list in that Annex, could suggest that in principle all criteria may be considered together in determining whether a model meets Article 51(1)(b)’s requirements. The force of this argument is, however, weakened by the observation that it appears possible to identify a division whereby the criteria in points (a) to (e) of Annex XIII primarily relate rather to a model’s capabilities, while the criteria in points (f) and (g) primarily relate to a model’s impact.²⁷³

68Moreover, the relevance of cumulative equivalence under Article 51(1)(b) finds some (albeit limited) support in Recital 111’s twelfth sentence, which states that designation decisions under Article 51(1)(b) should be taken ‘on the basis of an overall assessment’ of the Annex XIII criteria. In light of this recital, one could argue that Article 51(1)(b)’s main substantive requirements are not capabilities or impact as such but rather equivalence to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . However, as laid out above,²⁷⁴ there are grounds for not giving the recital’s reference to an ‘overall assessment’ too much weight as recitals may clarify the legislature’s intention but do not have binding legal force²⁷⁵ and a requirement of an overall assessment does not appear in the text of Article 51(1)(b) itself.²⁷⁶

69Having established that Article 51(1)(b) sets out substantive requirements that are distinct from Article 51(1)(a) and having analysed these distinct requirements, two related questions arise concerning the interaction between Article 51(1)(a) and (b): first, is it sufficient for classification under Article 51(1)(b) that the model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ?²⁷⁷ Second, can a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. prevent designation under Article 51(1)(b) by proving its model lacks high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ?²⁷⁸ Both questions turn on how Article 51(1)(b) operates in cases where the presence or absence of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. has been definitively established.

70Article 51(1)(b)’s wording suggests that a GPAI model with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. meets the substantive requirements for classification under Article 51(1)(b), as high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. are by definition equivalent to themselves.²⁷⁹ A systematic argument can, however, be made against designation of a GPAI model with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. on the basis of Article 51(1)(b) in conjunction with Article 52(4)’s first subparagraph. Such models are automatically classified as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. under Article 51(1)(a).²⁸⁰ Moreover, an instance where designation of such automatically classified models may be valuable – namely, to make a legally binding determination that a model meets Article 51(1)(a)’s classification condition in the absence of provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. notification pursuant to Article 52(1)’s first sentence – is arguably already addressed by Article 52(1)’s third sentence establishing the Commission’s power to designate GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .²⁸¹ In light of this provision, it would appear redundant to allow the Commission to designate a GPAI model with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. on the basis of Article 51(1)(b) in conjunction with Article 52(4)’s first subparagraph. The legislative choice to limit the scope of Article 52(5)’s procedure to contest designation to designation pursuant to Article 52(4)’s first subparagraph additionally suggests that the Commission may not use its powers under Article 52’s designation provisions interchangeably.²⁸²

71Article 3(65) characterises systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. as ‘specific to the high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. of general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’. This raises the question of whether a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. can prevent its model’s classification under Article 51(1)(b) by proving that it does not have high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . Under a literal reading, ‘specific to’ can either mean exclusive to – implying that only GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. can present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. – or characteristic of – implying that GPAI models with high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. typically present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. without excluding that GPAI models without such capabilities may under certain circumstances present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. as well. Arguments for and against each interpretation are discussed elsewhere.²⁸³ For the present discussion, it is interesting to consider the implications of different interpretations of Article 3(65)’s definition of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .

72Where ‘specific to’ is read as exclusive to, the classification of a GPAI model without high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. under Article 51(1)(b) appears hard to justify, as the model would – by definition – not present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . Where it is read as characteristic of, however, the possibility to classify a GPAI model without high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. under Article 51(1)(b) could not only be justified but even warranted. In that case, if Article 51(1)(b) would not allow for classification of GPAI models without high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. that nonetheless present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , these models would escape classification, given that Article 51(1)(a) permits only high-impact capabilities-based classification and the existence of further classification pathways outside of Article 51(1) appears at least very uncertain.²⁸⁴ It appears doubtful whether the existence of GPAI models presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. that cannot be classified as such can be reconciled with the AI Act’s rules for GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , as that would mean that these systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. could present themselves on the Union market without thorough assessment or mitigation.²⁸⁵

2.1.2.2. Commission designation in the context of Article 51(1)(b)

73Under Article 51(1)(b) – in contrast to Article 51(1)(a)²⁸⁶ – models are not classified as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. automatically. Rather, classification under Article 51(1)(b) is ‘based on a decision of the Commission’ according to which the GPAI model has capabilities or an impact equivalent to those set out in Article 51(1)(a), that is, high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . According to Recital 111’s twelfth sentence, this decision should be taken ‘on the basis of an overall assessment’ (emphasis added) of the criteria listed in Annex XIII,²⁸⁷ a requirement which does not appear in the text of Article 51(1)(b) itself.²⁸⁸ Absent a Commission decision, models that meet Article 51(1)(b)’s substantive requirements are not considered GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .

74Although Article 51(1)(b) contains no express reference to Article 52(4)’s first subparagraph, compelling reasons suggest that the Commission decision referred to under Article 51(1)(b) constitutes a designation decision in the sense of Article 52(4)’s first subparagraph.²⁸⁹ Both provisions’ shared reference to Annex XIII as well as Recital 111’s characterisation of the Article 51(1)(b) decision as a designation decision²⁹⁰ strongly support this reading.²⁹¹

75The procedural rights laid down in Article 18 MSR apply mutatis mutandis for a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. facing designation of its model under Article 52(4)’s first subparagraph,²⁹² at least from 2 August 2026.²⁹³ In particular, the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. has the right to be heard under Article 18(3) MSR.²⁹⁴

2.1.3. Further classification pathways

76Intuitively, and by the article’s title ‘Classification rules’,²⁹⁵ Article 51(1) seems to exhaustively list all conditions under which a GPAI model may be classified as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . This section examines whether the AI Act, and particularly Article 52(1)’s third sentence, captures instances in which systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. materialise in a GPAI model with neither high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. nor equivalent capabilities or impact, thereby establishing alternative pathways to classification independent of those provided by Article 51(1). Convincing arguments speak against the existence of such alternative classification pathways beyond Article 51(1).²⁹⁶ Yet, in the subsequent Article 52 on ‘Procedure’, there are two provisions that establish the Commission’s power to designate GPAI models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. without directly referring to Article 51(1)(a) or (b) and their respective requirements.²⁹⁷ As such, neither Article 52(1)’s third sentence nor Article 52(4)’s first subparagraph expressly require that a GPAI model has ‘ high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ’ under Article 51(1)(a) or ‘capabilities or an impact equivalent to [ high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ]’ under Article 51(1)(b) for its designation as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .

77One can make sense of this absence of an express reference to Article 51(1)’s requirements in broadly two ways. First, this omission could be unintentional, meaning that these designation provisions do implicitly rely on the requirements for classification under Article 51(1)(a) and (b).²⁹⁸ Alternatively, however, the lack of a reference to Article 51(1) could have been deliberate – and thus of significance – implying that these designation provisions would establish independent pathways for designation of a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . Of course, more nuanced variations of these two broad interpretations are conceivable as well.²⁹⁹

78Under the first interpretation – that of Article 52 implicitly referring back to Article 51 – Article 52(1)’s third sentence would relate to Article 51(1)(a)’s high-impact capabilities-based classification, while Article 52(4)’s first subparagraph would correspond to classification under Article 51(1)(b), which is based on capabilities or impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .³⁰⁰ The arguments that support this reading are explored more extensively elsewhere.³⁰¹ In short, the argument mainly relies on the positioning of Article 52(1)’s third sentence immediately after the provisions on the notification obligation clearly relating to Article 51(1)(a), and the broad congruence of the wording of Article 52(4)’s first subparagraph and Article 51(1)(b).³⁰² In line with this interpretation, it can be argued that designation under Article 52(1)’s third sentence requires the GPAI model to have high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ³⁰³ and that designation under Article 52(4)’s first subparagraph requires the GPAI model to have capabilities or an impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .³⁰⁴ This interpretation is indeed convincing and shared by the Commission Guidelines.³⁰⁵

79However, the alternative interpretation – that these designation provisions establish independent classification pathways – merits consideration, particularly regarding Article 52(1)’s third sentence.³⁰⁶ This provision requires the Commission to become aware of ‘a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ’ without reference to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .³⁰⁷ Arguments centring on this wording support reading this provision as requiring the model to present risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. that fall under the definition of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. under Article 3(65).³⁰⁸

80Such additional classification pathways could play a role in the classification framework’s adaptability to evolving technological developments.³⁰⁹ Article 51(1)’s framework relies heavily on high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. as a proxy for systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . This particularly applies to classification under Article 51(1)(a), which directly requires such high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. . It applies to a lesser extent to classification under Article 51(1)(b) as well, which requires capabilities or an impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .³¹⁰ This approach fits with how the AI Act links the notion of systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. on a definitional level to the concept of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , defining it as ‘specific to the high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. of general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’.³¹¹ However, it may be less appropriate on the level of classification. There is considerable uncertainty regarding future technological developments and the ability of evaluation methods to keep up with the pace of capability improvements. Presently, the determination of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. relies heavily on the training compute threshold in Article 51(2),³¹² but training compute could become a less meaningful proxy for capabilities as AI development methods evolve.³¹³ Where one cannot reliably evaluate whether a model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , having classification pathways that function independently from capability assessments would be crucial. This could particularly be the case in scenarios where clear evidence that systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. have already materialised is available – for example, because of the occurrence of serious incidents Article 3(49) AI Act: ‘serious incident’ means an incident or malfunctioning of an AI system that directly or indirectly leads to any of the following: (a) the death of a person, or serious harm to a person’s health; (b) a serious and irreversible disruption of the management or operation of critical infrastructure; (c) the infringement of obligations under Union law intended to protect fundamental rights; (d) serious harm to property or the environment. – but strong approximations for a model’s capabilities are not available.³¹⁴ Besides impact-based classification under Article 51(1)(b),³¹⁵ an additional systemic risk-based classification pathway under Article 52(1)’s third sentence could fulfil this function.

81Strong arguments exist against such additional classification pathways, however. Although Article 51(1)’s wording does not expressly exclude the existence of further classification pathways,³¹⁶ it provides the main argument against them. If the drafters had intended to create pathways beyond Article 51(1)(a) and (b), the logical place for such provisions would have been within the classification provision itself. Furthermore, impact-based classification under Article 51(1)(b) may already permit classification based on actual or reasonably foreseeable negative effects of a model,³¹⁷ which would render an additional classification pathway largely redundant.

2.1.4. Effects of classification

82A model’s classification under Section 1. of Chapter V has the effect that the AI Act’s provisions for GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. apply.³¹⁸ The legislature presumably considered this so apparent that it refrained from specifically providing for it.³¹⁹ Nevertheless, this consequence follows from a number of textual indicators, including the positioning of the classification rules at the beginning of Chapter V of the AI Act, the correspondence in wording between Article 51(1) and Article 55(1),³²⁰ and Article 52(5) and (3) implying that classification leads to the GPAI model being ‘considered to present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ’.

83Most notably, providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. face obligations through Article 55(1), requiring them to perform model evaluations, assess and mitigate systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. associated with the model, report serious incidents Article 3(49) AI Act: ‘serious incident’ means an incident or malfunctioning of an AI system that directly or indirectly leads to any of the following: (a) the death of a person, or serious harm to a person’s health; (b) a serious and irreversible disruption of the management or operation of critical infrastructure; (c) the infringement of obligations under Union law intended to protect fundamental rights; (d) serious harm to property or the environment. and ensure an adequate level of cybersecurity protection.³²¹ Beyond these obligations, special provisions for GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. and their providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. appear throughout Section 2. (‘Obligations for Providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of General-Purpose AI Models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’) of Chapter V, comprising Articles 53 and 54, together with the related Annex XI. Under Article 53(1)(a) in conjunction with the second section of Annex XI, providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. must draw up, keep up-to-date and provide, upon request, to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and the national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. , additional information relating to the technical documentation of the model.³²² Furthermore, under Article 53(2)’s second sentence and Article 54(6), providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. do not benefit from exceptions for certain open-source GPAI models from relevant obligations under Articles 53 and 54.³²³ Article 52(6) additionally provides for the publication of a list of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .³²⁴

84Provisions specifically concerning GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. extend beyond Chapter V of the AI Act to Section 5 (‘Supervision, Investigation, Enforcement and Monitoring in Respect of Providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of General-Purpose AI Models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’) of Chapter IX and Chapter XII (‘Penalties’) of the AI Act. Under Article 92(1)(b), the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. may conduct evaluations of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. to investigate systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. at Union level;³²⁵ under Article 93(1)(b), the Commission may request a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to implement mitigation measures where an evaluation carried out in accordance with Article 92 has given rise to serious and substantiated concern of a systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. at Union level;³²⁶ and under Article 101(1)(d), a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of a GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. may be fined where it failed to make available to the Commission access to the model with a view to conducting such evaluations.³²⁷ Beyond these provisions specific to GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , the provisions for all GPAI models continue to apply to GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .³²⁸

85One may note that for the purpose of the abovementioned provisions, the classification rules under Section 1. of Chapter V establish the term general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. and its variants³²⁹ as a technical term whose meaning derives entirely from the classification process itself.³³⁰ Where substantive provisions apply to GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , the applicability of those provisions does not depend on an independent test of whether the GPAI model actually presents risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. that fall under the systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. definition of Article 3(65).³³¹ Any challenge on these grounds must in principle target the model’s classification itself through the procedure for contesting classification under Article 52(2) and (3), the procedure for contesting classification under Article 52(5) or an action for annulment before the Court of Justice of the European Union under Article 263(4) TFEU.³³² To permit otherwise would circumvent the specific requirements of these procedures and render systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. classification of any model perpetually open to challenge, which – as these very procedures demonstrate³³³ – contradicts the purpose of classification of GPAI models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . Conversely, the Commission cannot enforce substantive obligations specific to GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. without a model’s classification, requiring it to designate a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. where it is not automatically classified under Article 51(1)(a).

86Unlike the provisions for gatekeeper designation under the DMA and for designation of very large online platforms (VLOPs) and very large online search engines (VLOSEs) under the DSA,³³⁴ the AI Act does not provide for a transitional period between classification – either via automatic classification or through a designation decision – and the applicability of the obligations for GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . A provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. facing the prospect of classification of its model will therefore likely need to prepare to ensure compliance upon designation or automatic classification. Recital 112 clarifies in this context that a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should be able to foresee its model’s training compute threshold-based classification,³³⁵ which would allow it to prepare accordingly.

2.2. Article 51(2): Presumption of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models.

87Article 51(2) establishes a presumption of a GPAI model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. based on the cumulative amount of computation used for its training (“training compute”), as it is ‘[a]ccording to the state of the art at the time of entry into force of this Regulation […] one of the relevant approximations for model capabilities’.³³⁶

88Under the enacted version of the AI Act, the presumption is triggered when training compute exceeds 10²⁵ FLOPs,³³⁷ a threshold that some authors claim ‘is not based on empirical evidence but rather the result of a political compromise’.³³⁸ According to a June 2025 estimate, ‘over 30 publicly announced AI models from different AI developers’ surpassed this threshold.³³⁹ Given that training compute for frontier AI models has been estimated to have grown by more than four times per year between 2018 and 2024,³⁴⁰ updates to Article 51(2) appear reasonable to expect,³⁴¹ and Article 51(3) empowers the Commission to update the training compute threshold accordingly.³⁴²

89A GPAI model that meets this threshold is automatically classified as a GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. under Article 51(1)(a).³⁴³ Article 51(2) draws no distinction based on when during the development process the threshold is met, which may occur at an early training stage.³⁴⁴ Moreover, reaching the training compute threshold will likely be the primary trigger for the notification obligation under Article 52(1)’s first sentence.³⁴⁵

90Compute thresholds are not only relevant to the classification of GPAI models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . The Commission Guidelines envisage the use of a lower compute threshold to determine whether an AI model can be considered a GPAI model according to the definition under Article 3(63).³⁴⁶ They indicate that, if a model is trained with more than 1023 FLOPs and can generate language (either textually or through audio), images from text, or videos from text, this indicates that it can be considered a GPAI model.³⁴⁷ Training compute thresholds have also been explored as a means of AI regulation in the United States.³⁴⁸

2.2.1. Cumulative amount of computation used for training

91Article 51(2)’s presumption applies where ‘the cumulative amount of computation used for [the GPAI model’s] training’ exceeds 10²⁵ FLOPs. It is clear from this wording (‘training’) that the computation used for the model’s deployment, so-called inference compute,³⁴⁹ does not play a role for the presumption of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .³⁵⁰ However, as different activities and methods involving computation (“computational activities”) can play a role in the development of GPAI models,³⁵¹ a key question is which kinds of computational activities may be included in Article 51(2)’s compute count.

2.2.1.1. General rule

92The AI Act does not define the training of a GPAI model. However, Recital 111 sets out that ‘[t]he cumulative amount of computation used for training includes the computation used across the activities and methods that are intended to enhance the capabilities of the model prior to deployment, such as pre-training, synthetic data generation and fine-tuning’.³⁵² Drawing on this recital, the Commission Guidelines and some legal scholars have taken the view that, ‘as a general rule, […] all [computational activities] that contributed or will contribute to the model’s capabilities’ need to be taken into account.³⁵³ The Commission Guidelines specifically mention that computational activities directly contributing to parameter updates should be included.³⁵⁴

93A 2025 report prepared for the Commission’s Joint Research Centre proposed another approach to whether a computational activity is included in Article 51(2)’s training compute count based on two criteria.³⁵⁵ It proposed to take into account only computational activities that either directly update the final model’s parameters or create model-specific inputs that depend on the current model state.³⁵⁶ This approach has its merits since – as pointed out by the report – direct parameter updates ‘form the core of model training and are unambiguously responsible for enhancing model capabilities’.³⁵⁷ It is questionable, however, whether this consideration suffices for generally³⁵⁸ excluding computational activities that influence the final model’s weights only indirectly and do not create model-specific inputs. The distinction drawn between core and non-core computational activities finds no basis in the wording of Article 51(2), which does not distinguish between different computational activities performed for a model’s training and therefore suggests that all such computational activities need to be taken into account. Moreover, Article 51(2) refers to the ‘cumulative amount of computation used for [the GPAI model’s] training’ (emphasis added). While this could be an acknowledgement of the evident fact that compute expenditure needs to be accumulated over time, it may also suggest a broader conception of training compute.³⁵⁹ The same applies to Recital 111, which expressly mentions synthetic data generation as a computational activity that does not lead to direct parameter updates.³⁶⁰ Moreover, there is no generally agreed upon definition of an AI model’s training that excludes computational activities indirectly updating the final model’s parameters.³⁶¹ Introducing an additional criterion such as whether the computational activity leads to direct parameter updates may come with an increased risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. of regulatory arbitrage where providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may seek to circumvent Article 51(2) by interposing an additional step between a compute-intense activity and the parameter updates.³⁶²

94Similarly, the mere fact that model-dependent processes ‘generate training signals […] specifically tailored to the model’s capabilities or are inseparable from the capability development process’ (emphasis added) and ‘cannot be trivially reused for other training pipelines’ – the rationale presented for including model-dependent processes³⁶³ – does not appear sufficient justification for excluding model-independent processes that enhance a model’s capabilities from Article 51(2)’s compute count.³⁶⁴ Article 51(2)’s text does not differentiate between model-dependent and model-independent processes. Recital 111 specifies that Article 51(2) includes computational activities that are ‘intended to enhance the capabilities of the model prior to deployment’ (emphasis added). While this criterion could be interpreted in a way that comes close to the concept of ‘specifically tailored’ training signals,³⁶⁵ Recital 111 simultaneously cites synthetic data generation as an example of a computational activity that need not be model-dependent but should nevertheless be included in the cumulative amount of compute used for training under Article 51(2).³⁶⁶ Moreover, the criterion of model-dependency may not be able to fully account for the increasing complexity of model training pipelines where different models may play a role in the process of training a GPAI model.³⁶⁷ For example, where training includes the training of different models that are later combined into one model,³⁶⁸ it may be unclear to which model(s) the model-dependency criterion applies. It is also uncertain how the criterion relates to the AI Act’s understanding of a model’s lifecycle,³⁶⁹ which may imply that different models used in training – such as parent models used in distillation³⁷⁰ – coincide with the final model from the AI Act’s legal point of view.³⁷¹ Finally, the introduction of this additional criterion could give rise to a greater risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. of gamification as well.

2.2.1.2. Exceptions

95Proceeding on the basis of the general rule that, in principle, all computational activities contributing to the model’s capabilities are to be included in Article 51(2)’s training compute count,³⁷² the question arises of whether exceptions exist to this general rule. The Commission Guidelines suggest the existence of such exceptions and list six examples of computational activities that, seemingly in contrast to the general rule,³⁷³ do not need to be counted towards Article 51(2)’s training compute threshold, while highlighting that this list may be subject to change due to technological developments.³⁷⁴ The listed examples are (i) the generation of publicly accessible synthetic data,³⁷⁵ (ii) ‘purely diagnostic processes’ not contributing to the model’s capabilities (such as model evaluations), (iii) computational activities which ‘contribut[e] to enhancing model capabilities only through lessons learnt by humans’ (such as exploratory research projects or failed experiments in synthetic data generation),³⁷⁶ (iv) the training of parent models used in distillation,³⁷⁷ (v) the training of auxiliary models (such as reward models) and (vi) recomputation of activations to save memory.³⁷⁸

96It is interesting to consider the extent to which, and the grounds upon which, the exclusion of these examples and other potential exceptions can be justified. The second and sixth examples – purely diagnostic processes and recomputations of activations – appear most straightforward to justify, as they rather represent a confirmation of the general rule. Purely diagnostic processes, by definition, do not contribute to the model’s capabilities.³⁷⁹ Recomputation of activations has been developed as a technique to address memory limitations of the hardware used in AI model training.³⁸⁰ To avoid failure of model training due to running out of memory, some of the activations, which may be described as intermediate computational results generated during the model’s training,³⁸¹ are not stored but discarded and later regenerated.³⁸² As these recomputations in essence ‘trad[e] less memory for greater computation’,³⁸³ they can also be regarded as not resulting in an enhancement of a model’s capabilities.³⁸⁴

97The other examples listed in the Commission Guidelines appear less straightforward on the basis of the Act’s text. In particular, special attention must be paid to the particular cases of synthetic data generation and knowledge distillation, which are discussed in more detail in subsequent sections.³⁸⁵ In general, when considering the scope of what constitutes ‘cumulative training compute’, it is necessary to consider Article 51(2)’s integration within the systematic context of the classification rules in Section 1. of Chapter V. Many concerns that could, in principle, support an exception for certain types of compute appear to be already addressed to some extent by the AI Act’s provision for a procedure to contest the training compute-based classification under Article 52(2) and (3), which allows providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to rebut Article 51(2)’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. presumption.³⁸⁶ This applies, for example, in cases where a computational activity only marginally contributes to the enhancement of a model’s capabilities or has a particularly low compute-to-capability ratio, as such considerations could serve to rebut the high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. presumption. Moreover, far-reaching exceptions from the training compute count may also prove unnecessary in light of Article 51(3), which permits the Commission to adjust the compute threshold, including the computational activities counting towards the threshold, via a delegated act where necessary.³⁸⁷

98Synthetic data – which is not directly based on observation of the real world but rather artificially generated from real data³⁸⁸ – plays an increasingly important role in the training of AI models.³⁸⁹ Microsoft’s phi-4 model exemplifies this trend.³⁹⁰ Around twenty-five percent of the model’s pre-training compute budget was spent on the generation of synthetic data.³⁹¹ While only having 14 billion parameters and therefore being small in comparison with frontier AI models, phi-4 scored better on the mathematics and science benchmarks MATH Level 5 and GPQA Diamond than did OpenAI’s GPT-4o – the model that was used to generate the synthetic data for phi-4’s training.³⁹² The Commission Guidelines specify that the generation of non-publicly accessible synthetic data should be included in Article 51(2)’s compute count,³⁹³ while mentioning the generation of publicly accessible synthetic data as an example of a computational activity that should be excluded because it ‘may be indistinguishable from other publicly accessible data’.³⁹⁴ Some authors have even argued that the generation of synthetic data should not be included regardless of whether it is publicly available, contending in particular that synthetic data generation precedes rather than forms part of the model’s training.³⁹⁵

99Several arguments support these views. While no apparent consensus has emerged as to a definition of a GPAI model’s training, there are definitions of model training which appear to exclude synthetic data generation.³⁹⁶ Further, synthetic data serves as a functional replacement of real data whose generation is also not included in Article 51(2)’s compute count.³⁹⁷ Moreover, it appears possible that in certain instances – particularly where a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. acquires synthetic data from a third party – it is difficult to know whether the data has been synthetically generated and the amount of compute spent to generate the data.³⁹⁸

100However, there are also good reasons to include synthetic data generation.³⁹⁹ The inclusion of compute spent on synthetic data generation in a model’s pre-training compute budget by some authors suggests that including synthetic data generation within the notion of training does not exceed the wording of Article 51(2).⁴⁰⁰ This is reinforced by Recital 111 specifically mentioning synthetic data generation among the activities and methods to be included in Article 51(2)’s compute count, regardless of its public availability.⁴⁰¹

101Furthermore, in many cases the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should be able to estimate the amount of compute spent on synthetic data generation – for instance, by tracking compute expenditure where it generates the data itself or through contractual arrangements governing data purchases.⁴⁰² Indeed, Article 53(1)(a) requires providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to possess at least some information about synthetic data generation, which could serve as the basis for such estimates. This provision requires providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models to draw up technical documentation – including information on the data used for training, in particular the type and provenance of data and how it was obtained, as well as information on the computational resources used to train the model⁴⁰³ – for the purpose of providing it, upon request, to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and the national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. .⁴⁰⁴ This includes information on the use and provenance of synthetic data for the model’s training.⁴⁰⁵

102This documentation requirement may limit the risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. of regulatory arbitrage associated with including compute spent on synthetic data generation in Article 51(2)’s compute count, as providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. that seek to generate or acquire synthetic data in a manner that makes it impossible for them to estimate the compute spent on its generation may risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. breaching their obligation under Article 53(1)(a).⁴⁰⁶ Accordingly, cases where it proves genuinely impossible for the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to produce even an estimate of the compute spent on synthetic data generation may be edge cases that do not necessarily warrant the general exclusion of synthetic data generation.

103Overall, the more compelling arguments support including synthetic data generation in Article 51(2)’s compute count, in closer alignment with the Commission Guidelines’ general rule to include all computational activities that contribute to the model’s capabilities.⁴⁰⁷ This approach is further supported by the consideration that an additional differentiation – for example, between publicly and non-publicly accessible synthetic data⁴⁰⁸ – raises new demarcation questions, such as whether and under which circumstances it suffices for a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to offer synthetically generated data for sale in order for it to be considered publicly accessible. Moreover, cases where the compute threshold is surpassed due to synthetic data generation without significant capability enhancement can be addressed through the procedure to contest classification set out in Article 52(2) and (3).⁴⁰⁹

104(Knowledge) distillation is a method for model training that involves a smaller “student model” and a larger “teacher model”.⁴¹⁰ Rather than undergoing standard pre-training, the student model is trained to “mimic” the output of the teacher model.⁴¹¹ By “reusing” the teacher model,⁴¹² smaller student models require fewer computational resources and operate faster than their teacher models, without necessarily sacrificing capabilities.⁴¹³ For example, researchers successfully distilled Google’s BERT model into the DistilBERT model, using only 3% of the computational resources required for training the original model while retaining 97% of BERT’s language understanding capabilities and achieving 60% faster performance.⁴¹⁴

105Three computational activities associated with distillation warrant consideration under Article 51(2): first, the training of the teacher model; second, the generation of the teacher model’s outputs that serve as a basis for the training of the student model; and thirdly, the training of the student model with this output.⁴¹⁵ As with synthetic data generation, interpretive uncertainty arises as to whether some or all of these computational activities must be taken into account.

106Under the Commission Guidelines’ general rule laid out above,⁴¹⁶ all three computational activities would be included in the student model’s compute count under Article 51(2) as they all, at least indirectly, contribute to the student model’s capabilities.⁴¹⁷ However, there are arguments for an exception from this rule, in particular with regard to the first computational activity, the training of the teacher model.⁴¹⁸ One could contend that the amount of computation used for training a student model’s teacher model is not ‘used for its [that is, the student model’s] training’ (emphasis added) in the sense of Article 51(2). Moreover, where a teacher model serves purposes beyond distillation, one may argue that the training of the teacher model is not ‘intended to enhance the capabilities of the [student] model’ in the sense of Recital 111’s sixth sentence but rather serves the training of the teacher model itself.⁴¹⁹

107While these arguments carry some weight, they do not appear entirely convincing. As evidenced by Recital 111’s reference to synthetic data generation, there are good reasons to assume that a computational activity is not required to directly contribute to the (student) model’s capabilities to be included in Article 51(2)’s compute count.⁴²⁰ Additionally, it is questionable whether the fact that a computational activity serves an additional purpose distinct from contributing to the (student) model’s capabilities is a sufficient reason for exclusion. By its wording, Article 51(2) does not require a computational activity to only serve the model’s training in order to be included. This is reinforced by a purposive argument that such an exclusivity requirement would lead to greater delineation difficulties and loopholes as a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. could seek to avoid the inclusion of a computational activity by assigning a secondary purpose to it.⁴²¹

108Considering whether and under which circumstances the teacher model and the student model may be regarded as one and the same model under the AI Act adds an additional layer of complexity to the question of whether the teacher model’s training compute should be included in the student model’s compute count. There appears to be no simple answer to the relationship between a teacher model and a student model under the AI Act, which relates to the broader question of how the AI Act treats modification and subsequent development of a model.⁴²² The Commission Guidelines assume that where the same provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. provides both the teacher model and the student model, ‘[a]ny subsequent development of the model downstream of [the model’s] large pre-training run […] forms part of the same model’s lifecycle rather than giving rise to new models’ while also acknowledging the difficulty of delineating a model and its lifecycle.⁴²³ If one were to follow this and further regard distillation as a ‘subsequent development of the model’, a teacher model and its student model could be one and the same under the AI Act, likely implying that the teacher model’s training compute would need to be taken into account. A purposive argument in favour of this interpretation is that it could avoid a potential loophole described in literature: a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. could develop a teacher model above the training compute threshold, use distillation to train a student model below the threshold, and only place the student model on the market.⁴²⁴ However, such an interpretation appears far from certain, and the Commission Guidelines’ approach only relates to the case where the same provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. provides both the original and modified model.⁴²⁵

109To further add to the complexity, one can inquire about which kind of allocation of responsibilities between a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of a teacher model and a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of a student model with regard to the obligations under Article 55 stemming from systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ⁴²⁶ classification would be most faithful to the purpose of these obligations. It appears that certain systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. can be most effectively, or even only, addressed by the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the teacher model, while others are better addressed by the student model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. .⁴²⁷ Moreover, a categorical focus on only one of the providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. could be insufficient for the effective mitigation of systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. where one of the models is not placed on the Union market and is therefore outside the scope of the AI Act.⁴²⁸ Conversely, arguments against unnecessary double obligations for both the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the student model and the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the teacher model can be made, at least where this does not contribute to a more effective mitigation of systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .⁴²⁹

2.2.1.3. Methods for determining the cumulative amount of computation

110As there are different computational activities that may contribute to a model’s cumulative amount of computation,⁴³⁰ there are different conceivable methods for determining the amount of computation spent on each of these computational activities.⁴³¹ This observation is reflected in point (c) of Annex XIII, which lays out that the amount of computation used for a model’s training can be both ‘measured in floating point operations’ or ‘indicated by a combination of other variables’, including estimates of training cost, duration or energy consumption.⁴³²

111The Commission Guidelines state that a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ‘may choose any method to estimate the relevant amount of training compute, so long as the estimated amount is, in the providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’ best judgement, accurate within an overall error margin of 30% of the reported estimate.’⁴³³ The use of some error margin appears appropriate given the difficulties in precisely measuring compute expenditure across all relevant computational activities.⁴³⁴ However, a more dynamic permissible overall error margin instead of relying on a fixed overall error margin of 30% may be more appropriate in the context of Article 51(2). While a fixed error margin appears particularly sensible where a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must determine computational resources independently of any specific threshold,⁴³⁵ it may impose unnecessary estimation effort and produce difficult-to-justify results regarding the question of whether the training compute threshold has been crossed. Two examples may illustrate these difficulties:

112First, suppose a method estimates a model’s cumulative amount of computation at 1024 FLOPs with an overall error margin of 50%. This would mean that the model was trained with no more than 1.5 × 1024 FLOPs, compellingly proving it does not meet Article 51(2)’s 10²⁵ FLOPs threshold. Yet the Commission Guidelines would deem this method impermissible because its error margin exceeds 30%. Second, suppose a method estimates a model’s cumulative amount of computation to be 10²⁵ FLOPs with an overall error margin of 30%. Depending on the positioning of the estimate within this error margin, there could be a substantial probability that the model exceeds the threshold. Yet the Commission Guidelines suggest this method would constitute sufficient proof that the model does not exceed Article 51(2)’s threshold.

113These examples illustrate that setting a permissible overall error margin for reported estimates in the abstract may produce undesirable results. Moreover, it appears difficult to derive such a threshold from the AI Act itself, which provides no guidance on permissible error margins. Rather, a different approach, relying on a more dynamic error margin, appears worth considering: the closer a model’s estimated computation is to Article 51(2)’s threshold, the greater the precision required of a method to prove whether the threshold is met or not.⁴³⁶ Such an approach would come with its own limitations⁴³⁷ but could be suitable to avoid the difficult-to-justify results laid out above.

114An additional suggestion by some scholars that computational activities consuming less than 20% of the overall compute budget cumulatively should not be taken into account also has some practical appeal.⁴³⁸ In many cases, activities accounting for only a small portion of the training compute budget may indeed be negligible. However, the proposal may lack necessary nuance where a model’s training compute is close to the set threshold, as these computational activities could prove decisive in determining whether the training compute threshold is met.

115The Commission Guidelines lay out two possible approaches for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to estimate training compute – a hardware-based approach and an architecture-based approach.⁴³⁹ Under the hardware-based approach, the (expected) training compute is estimated based on the number of graphics processing units (“GPUs”) or other hardware units used for training, the total duration of their use, their peak theoretical performance and their average percentage of utilisation.⁴⁴⁰ This differs from the architecture-based approach, which estimates the (expected) training compute based on the number of full passes made during the training of a neural network and the total number of operations performed in a full pass.⁴⁴¹ Alternatively, according to the Commission Guidelines, for some models based on a dense transformer architecture, the (expected) training compute may be estimated based on the total number of model parameters and the total number of training tokens used for training.⁴⁴²

116As the AI Act does not prescribe a particular method of determining whether a model exceeds Article 51(2)’s compute threshold, the approaches laid out in the Commission Guidelines appear merely illustrative, with other methods being admissible as well.⁴⁴³ The Commission Guidelines, by their nature, cannot bind GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. .⁴⁴⁴

117Moreover, an obligation to determine the cumulative amount of computation used for the model’s training via a ‘dual estimation methodology’ using both a hardware-based approach and the architecture-based approach, suggested by some scholars,⁴⁴⁵ does not appear supported by the AI Act or the Commission Guidelines.⁴⁴⁶ Further methods that could play a role in determining or verifying whether Article 51(2)’s threshold is met could include inferences from the cost, duration or energy consumption of training;⁴⁴⁷ inferences from benchmark performance;⁴⁴⁸ or testimony from whistleblowers.⁴⁴⁹

118All operations should be counted equally, independent of floating-point precision (e.g. FP8, FP16, FP32⁴⁵⁰).⁴⁵¹ By its wording, Article 51(2) is concerned with the number of FLOPs, not with the precision of the number format used in these operations. This interpretation is further reinforced by the purpose of Article 51(2)’s threshold, which serves as an indicator for a model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , since floating-point precision appears to be only marginally indicative of a model’s capabilities.⁴⁵²

119According to its Guidelines, the Commission expects providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ‘to document the assumptions made in making their estimations, including the method of estimation, and the associated uncertainties.’⁴⁵³ Such a documentation requirement implicitly follows from Article 52(1)’s second sentence as well. Where a model meets Article 51(2)’s training compute threshold, its provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. is not only required to notify the Commission of this fact but also to provide the information necessary to demonstrate that this is the case.⁴⁵⁴ While there is uncertainty as to the extent of information a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must submit, this includes information about the method of estimation.⁴⁵⁵

2.2.2. Effect of the presumption and rebuttability

120Where a model meets Article 51(2)’s training compute threshold, it is presumed to have high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. pursuant to Article 51(1)(a)⁴⁵⁶ and is thus automatically classified as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .⁴⁵⁷ Via Article 51(1)(a), this high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. presumption is further linked to the notification obligation under Article 52(1)’s first sentence.⁴⁵⁸ It is crucial to note that the notification obligation is triggered not only when the compute threshold is actually met, but also – as supported by various compelling arguments – when the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. already knows that it will be met.⁴⁵⁹

121Article 51(2)’s presumption is rebuttable.⁴⁶⁰ This is underscored by Article 52(2), which sets out a procedure for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to contest Article 51(1)(a)’s classification of GPAI models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. together with notification.⁴⁶¹ This procedure arguably allows the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to present arguments that, although the model meets or will meet the compute threshold, it does not have high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. and therefore should not be classified as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .⁴⁶²

122To rebut the presumption of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must prove that the model’s capabilities do not meet the definition of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. under Article 3(64) – that is, that they do not ‘match or exceed the capabilities recorded in the most advanced general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’.⁴⁶³ To establish this, the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. can make use of appropriate technical tools and methodologies referred to under Article 51(1)(a),⁴⁶⁴ with the Commission Guidelines specifically mentioning actual and forecasted benchmark results as potential grounds for rebuttal of the presumption.⁴⁶⁵ Moreover, the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may also argue that the fact that the model exceeds the threshold does not, in the particular circumstances, indicate high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .⁴⁶⁶ This could be the case where the cumulative amount of training compute barely surpasses the threshold set by Article 51(2)⁴⁶⁷ or where it is only met due to the inclusion of certain computational activities with a particularly low compute-to-capability ratio.

2.3. Article 51(3): Delegated acts

123Article 51(3) provides that ‘[t]he Commission shall adopt delegated acts in accordance with Article 97 to amend the thresholds listed in paragraphs 1 and 2 of this Article, as well as to supplement benchmarks and indicators in light of evolving technological developments, such as algorithmic improvements or increased hardware efficiency, when necessary, for these thresholds to reflect the state of the art’.⁴⁶⁸ This delegation of power to the Commission is based on Article 290(1) TFEU,⁴⁶⁹ and the powers to ‘amend’ and to ‘supplement’ correspond to the two distinct categories of delegated powers laid down in this provision.⁴⁷⁰ In general, a power to ‘amend’ a legislative act aims to ‘authorise the Commission to modify or repeal non-essential elements’ of an act, whereas a power to ‘supplement’ a legislative act aims to ‘authorise the Commission to flesh out that act’.⁴⁷¹ Like other provisions of the AI Act empowering the Commission to adopt delegated acts, Article 51(3) aims to allow for necessary updates to the regulatory framework.⁴⁷²

2.3.1. Scope of the delegation of power

124As laid out above, the AI Act refers to a multitude of assessment instruments in the context of classification, such as (technical) tools, methodologies, indicators, benchmarks, criteria, thresholds, approximations and evaluations, without defining them or clearly distinguishing between them – and potentially using some of them synonymously.⁴⁷³ This inconsistent terminology creates uncertainty regarding the scope of Article 51(3)’s delegation of power, which specifically mentions thresholds, benchmarks and indicators.

125As a result, some scholars have interpreted the scope of the delegation of power under Article 51(3) broadly, encompassing all relevant criteria under Article 51,⁴⁷⁴ while other scholars see the scope of Article 51(3) more narrowly, applicable in particular to Article 51(2)’s training compute threshold.⁴⁷⁵ The Commission Guidelines, staying close to Article 51(3)’s wording, set out that the provision empowers the Commission to ‘adjust the thresholds set out in Article 51(1) and (2) AI Act’ and ‘to introduce additional benchmarks and indicators’.⁴⁷⁶

126The subsequent sections examine the scope of Article 51(3)’s delegation of power with regard to (i) Article 51(2)’s training compute threshold;⁴⁷⁷ (ii) Article 51(1)’s substantive criteria for classification;⁴⁷⁸ (iii) indicators and benchmarks referred to under Article 51(1)(a);⁴⁷⁹ and (iv) the criteria contained within Annex XIII.⁴⁸⁰

2.3.1.1. Training compute threshold under Article 51(2)

127Based on the wording of Article 51(3) and the relevant recitals, it is clear that Article 51(3) empowers the Commission to amend Article 51(2)’s training compute threshold.⁴⁸¹ The examples of technological developments stated in Article 51(3) that could necessitate an update of this threshold – ‘algorithmic improvements’ and ‘increased hardware efficiency’ – suggest that the legislature primarily contemplated a lowering of the threshold. However, neither Article 51(3) nor the recitals exclude a raising of the threshold.⁴⁸²

2.3.1.2. Substantive criteria for classification under Article 51(1)

128Some scholars argue that Article 51(3) also empowers the Commission to amend the substantive criteria for classification listed under Article 51(1).⁴⁸³ Such a broad interpretation of Article 51(3) – that requires interpreting these substantive criteria as ‘thresholds listed in paragraphs 1 and 2’ in the sense of Article 51(3) – could have profound implications, as it would grant the Commission the power to introduce new conditions for the classification of a GPAI model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. besides a models’ high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ⁴⁸⁴, or its capabilities or impact equivalent to high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. ⁴⁸⁵. This, in turn, could lead to the classification of new or different models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. that may currently not be within scope of the classification conditions under Article 51(1).⁴⁸⁶

129This broad interpretation does not exceed the wording of Article 51(3), as a substantive criterion such as the model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. can be referred to as a qualitative threshold.⁴⁸⁷ Indeed, Recital 111 can be read in such a way.⁴⁸⁸ Moreover, this interpretation would explain why Article 51(3)’s power to amend thresholds does not refer only to the training compute threshold in Article 51’s second paragraph but to its first paragraph as well – which otherwise remains difficult to explain.

130In particular, it is not entirely convincing to interpret the power ‘to amend thresholds listed in paragrap[h] 1’ under Article 51(3) as referring to the assessment instruments for evaluating whether a model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. mentioned in Article 51(1)(a).⁴⁸⁹ Article 51(1)(a) does not mention thresholds but ‘technical tools and methodologies, including indicators and benchmarks’.⁴⁹⁰ Moreover, Article 51(3) refers to the power ‘to amend thresholds listed in paragrap[h] 1’ separate from the power ‘to supplement benchmarks and indicators’.⁴⁹¹ Since the latter refers to assessment instruments mentioned in Article 51(1)(a),⁴⁹² interpreting the former as referring to them as well is not readily apparent. Additionally, if the legislature had intended the power ‘to amend thresholds listed in paragrap[h] 1’ (emphasis added) to refer to Article 51(1)(a)’s assessment instruments, it would be difficult to explain why it granted the power to ‘amend’ rather than to ‘supplement’ thresholds in this context.⁴⁹³ The introduction of new thresholds as assessment instruments for evaluating a model’s high-impact would constitute – like the introduction of benchmarks and indicators – a matter of fleshing out Article 51(1)(a)’s provision on assessment instruments rather than modifying or repealing it.⁴⁹⁴ Nor is it entirely convincing to interpret the power ‘to amend the thresholds listed in paragrap[h] 1’ as referring to the business user threshold contained in point (f) of Annex XIII, since Article 51(1)(b) contains only a reference to this annex while the threshold itself is not contained in Article 51(1).⁴⁹⁵ In light of these considerations, an interpretation of Article 51(3) as empowering the Commission to amend the substantive criteria for classification listed under Article 51(1) could ensure that the provision for the Commission’s power ‘to amend thresholds listed in paragrap[h] 1’ retains its effectiveness.⁴⁹⁶

131However, it is questionable to what extent the power to determine the substantive requirements for classification of a GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. can be delegated. Article 290(1) TFEU requires that ‘[t]he essential elements of an area shall be reserved for the legislative act and accordingly shall not be the subject Article 3(58) AI Act: ‘subject’, for the purpose of real-world testing, means a natural person who participates in testing in real-world conditions. of a delegation of power’. Whether and to what extent the substantive requirements under Article 51(1) constitute ‘essential elements’ within the meaning of Article 290(1) TFEU appears highly uncertain.⁴⁹⁷

132Further, a purposive argument may be advanced against this broad interpretation of Article 51(3) as allowing for the amendment of Article 51(1)(a)’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. requirement. The legislature has already shaped the concept of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. in a way that is responsive to evolving technological developments by defining it in relation to the ‘capabilities recorded in the most advanced general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’.⁴⁹⁸ Consequently, it remains uncertain whether an amendment of the substantive criteria under Article 51(1) can be ‘necessary […] in light of evolving technological developments, such as algorithmic improvements or increased hardware efficiency’ in the sense of Article 51(3).⁴⁹⁹

133In case of an amendment of the substantive criteria for classification under Article 51(1)(b), this would, in principle, necessitate an amendment of Annex XIII’s introductory wording as it mirrors Article 51(1)(b)’s language.⁵⁰⁰ Interestingly, Article 52(4)’s second subparagraph by its wording only empowers the Commission to specify and update the criteria but not the chapeau of Annex XIII. Moreover, it is uncertain whether Article 51(3) itself allows for an amendment of Annex XIII.⁵⁰¹

2.3.1.3. Indicators and benchmarks under Article 51(1)(a)

134Article 51(1)(a) mentions not only the substantive criterion of a model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. but also ‘indicators and benchmarks’ for their evaluation. Article 51(3) empowers the Commission to introduce and specify these assessment instruments via a delegated act.⁵⁰²

2.3.1.4. Criteria contained within Annex XIII

135Whether and to what extent the delegation of power under Article 51(3) encompasses thresholds, benchmarks and indicators contained in Annex XIII, is unclear.⁵⁰³ This question matters: although Article 52(4)’s second subparagraph specifically empowers the Commission to amend Annex XIII by specifying and updating the criteria, that only establishes the Commission’s power to adopt delegated acts, whereas Article 51(3) arguably obliges the Commission to make use of the corresponding delegation of power under certain circumstances.⁵⁰⁴ Such an obligation could be particularly relevant with regard to point (f) of Annex XIII. This provision contains a quantitative threshold as it provides that a model’s high impact on the internal market due to its reach shall be presumed when it has been made available to at least 10,000 registered business users established in the Union.⁵⁰⁵ Thresholds and benchmarks are further expressly mentioned in points (d) and (e) of Annex XIII.⁵⁰⁶

136However, extending the delegation of power under Article 51(3) to thresholds contained in Annex XIII is not close at hand since this delegation provision specifically refers to thresholds listed in Article 51’s first and second paragraph.⁵⁰⁷ With regard to benchmarks and indicators contained in Annex XIII, it is difficult to envisage that their supplementation would ever be necessary in the sense of Article 51(3), particularly if one views the list of criteria as non-exhaustive.⁵⁰⁸ Moreover, Annex XIII refers to its content as ‘criteria’ rather than as thresholds, benchmarks or indicators.⁵⁰⁹ Since Article 51(3) – unlike Article 52(4)’s second subparagraph⁵¹⁰ – neither uses the term ‘criteria’ nor expressly references Annex XIII, this provides a further argument against the delegation of power under Article 51(3) encompassing criteria contained in Annex XIII.

2.3.2. Obligation to adopt delegated acts

137The wording of Article 51(3) (‘the Commission shall adopt delegated acts’) suggests an obligation of the Commission to make use of the powers delegated to it under certain instances.⁵¹¹ Such an obligation to adopt delegated acts, recognised under EU law more generally,⁵¹² is supported by comparison to the wording of other delegation provisions in Chapter V. Under Article 52(4)’s first subparagraph and Article 53(5) and (6), the Commission ‘is empowered to adopt delegated acts’ (emphasis added) to update Annex XIII and Annexes XI and XII, respectively. In general, the distinction between shall and is empowered is not merely of stylistic nature because shall generally imposes a binding obligation, while being empowered signals discretion that the Commission may exercise.⁵¹³

138The AI Act’s recitals support this distinction in the present case, which state with regard to Article 51(3) that the ‘threshold [of floating point operations under Article 51(2)] should be adjusted over time to reflect technological and industrial changes, such as algorithmic improvements or increased hardware efficiency, and should be supplemented with benchmarks and indicators for model capability’ (emphasis added).⁵¹⁴ This interpretation is in line with the Commission Guidelines which state with respect to the technical tools and methodologies mentioned under Article 51(1)(a) that ‘[t]hese tools and methodologies are to be further specified by the Commission through adoption of delegated acts […]’ (emphasis added).⁵¹⁵

139The Commission will likely enjoy considerable discretion in determining whether this obligation is triggered. The legislature has not tied the duty to exercise the delegated power to specific timeframes but made it contingent upon whether it is ‘necessary, for these thresholds to reflect the state of the art’. Nonetheless, there are indications that the legislature intended only a limited number of providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. to be classified at a time.⁵¹⁶ Article 265 TFEU provides for an action for failure to act before the Court of Justice of the European Union to have the infringement established.⁵¹⁷

140It is unlikely that a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. could challenge the classification of its model as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. on the grounds that the Commission breached its obligation to amend the threshold. Such reasoning would effectively make the threshold’s applicability contingent upon the Commission’s fulfilment of its obligation under Article 51(3), while the legislature did not expressly provide for Article 51(2)’s threshold’s conditional applicability.⁵¹⁸ Nor is there any indication that such conditional applicability of Article 51(2)’s threshold was intended. The example scenarios mentioned in Article 51(3) – ‘algorithmic improvements or increased hardware efficiency’ – suggest that the legislature primarily envisaged a lowering of the training compute threshold.⁵¹⁹ In such scenarios, the relevant concern would not be whether models were wrongly classified under an outdated threshold, but rather whether models escaped classification that nevertheless present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. despite having been trained with lower compute, for example because of algorithmic improvements. Moreover, meeting Article 51(2)’s threshold merely gives rise to a presumption of high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , which its provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. can rebut by contesting the model’s classification under Article 52(2)–(3).⁵²⁰ This contestation procedure allows a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to prevent its model’s classification based on a potentially outdated threshold, where the model does not present systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .⁵²¹ In particular, in the context of this contestation procedure, the Commission takes into account how indicative the model’s training compute is for assessing whether it has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. .⁵²²

2.3.3. Conditions for the adoption of delegated acts

141The power to adopt delegated acts under Article 51(3) is conferred on the Commission subject to the conditions laid down in Article 97.⁵²³ The delegation runs for five years from 1 August 2024, with tacit extension for periods of an identical duration absent opposition by the European Parliament and the Council no later than three months before the end of each period.⁵²⁴ Both institutions retain the power to revoke the delegation at any time.⁵²⁵ Before adopting any delegated act, the Commission must consult Member State experts designated by the Member States in line with the principles established in the Interinstitutional Agreement on Better Law-Making of 13 April 2016.⁵²⁶ Once adopted, the delegated act must be notified to the European Parliament and the Council,⁵²⁷ which then have three months (extendable to six months) to raise objections that would prevent the act from taking effect.⁵²⁸

2.4. Annex XIII

2.4.1. Overview

142Annex XIII contains a heterogeneous list of around eleven⁵²⁹ criteria with relevance for systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. classification of GPAI models.⁵³⁰ These criteria are particularly relevant for Commission designation decisions under Article 51(1)(b) in conjunction with Article 52(4)’s first subparagraph, where the Commission must take these criteria into account.⁵³¹ However, they may also play a role in other instances connected to the classification of GPAI models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , such as reassessment of a designation under Article 52(5) or the assessment of whether a GPAI model has high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. under Article 51(1)(a).⁵³²

143The criteria, which are grouped in seven points and already touched upon above,⁵³³ range from rather specific, technical and quantifiable (e.g. the number of model parameters⁵³⁴) to rather unspecific (e.g. the evaluations of capabilities of the model⁵³⁵), non-technical (e.g. the number of registered end users⁵³⁶) and not easily quantifiable (e.g. the quality of the data set⁵³⁷). Some criteria are more related to a model’s capabilities (e.g. the evaluations of the model’s capabilities⁵³⁸), whereas others are more related to its impact (e.g. the number of registered end users⁵³⁹).⁵⁴⁰ In some instances, Annex XIII only lists the criterion itself (e.g. the number of model parameters⁵⁴¹), while in others it provides illustration through examples (e.g. biological sequences as a specific type of model input⁵⁴²), indications of measurement units (e.g. FLOPs for amount of computation used for training⁵⁴³), and measurement methods (e.g. measuring the size of the data set through tokens⁵⁴⁴). Regarding the model’s high impact on the internal market due to its reach, point (f) of Annex XIII notably contains a presumption based on the number of registered business users.⁵⁴⁵

144Article 52(4)’s second subparagraph allows the Commission to ‘specif[y] and updat[e]’ the Annex XIII criteria.⁵⁴⁶ In this respect, the more compelling arguments appear to support the view that the Commission is also empowered to add new criteria to the list.⁵⁴⁷ Beyond this, it appears rather doubtful whether Article 51(3)’s delegation of power extends to Annex XIII as well.⁵⁴⁸ This could be relevant insofar as the Commission might, under certain conditions, be obliged to adapt the threshold of 10,000 business users under point (f) of Annex XIII to technological developments.⁵⁴⁹

2.4.2. Relevance for classification

145Annex XIII has overarching significance for the classification rules in Section 1. of Chapter V,⁵⁵⁰ which expressly mentions it four times. Article 51(1)(b) serves as Annex XIII’s primary anchor point,⁵⁵¹ as confirmed both by Annex XIII’s title referring to Article 51 and by the fact that Article 51(1)(b) is the only provision in Article 51 that expressly references the Annex. Annex XIII’s wording reinforces this connection by stating that it contains a list of criteria that the Commission shall take into account ‘[f]or the purpose of determining that a general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. has capabilities or an impact equivalent to those set out in Article 51(1), point (a)’ – language that mirrors Article 51(1)(b). The express reference to Article 51(1)(a) in Annex XIII should therefore not be misunderstood as indicating that Annex XIII relates directly to that provision⁵⁵² – rather, the reference reflects the fact that Article 51(1)(a) is itself expressly mentioned in Article 51(1)(b).

146Annex XIII’s relevance to classification is further reflected in Article 52 as well. Article 52(4)’s first subparagraph establishes the Commission’s power to designate GPAI models as presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ‘on the basis of criteria set out in Annex XIII’,⁵⁵³ while Article 52(5)’s first sentence clarifies that reassessment of such designations must likewise be made on the basis of Annex XIII’s criteria.⁵⁵⁴ The reference to Annex XIII in Article 52(4)’s first subparagraph provides additional support for the view taken here that the designation decision under that provision coincides with the Commission decision referred to in Article 51(1)(b).⁵⁵⁵

147While Annex XIII is not expressly mentioned in the provisions relating to the classification based on a model’s high-impact capabilities Article 3(64) AI Act: ‘high-impact capabilities’ means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. , its criteria may still be considered when deciding whether a model satisfies Article 51(1)(a) or whether, due to its specific characteristics, it does not present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. under Article 52(2) and (3).⁵⁵⁶ Several criteria in Annex XIII appear to serve as indicators of a model’s (high-impact) capabilities rather than its impact. Examples include the amount of computation used for the model’s training – mentioned not only in Article 51(2) but also in Annex XIII, point (c) – and the benchmarks and evaluations of model capabilities referred to in point (e) of the Annex.