Cambridge Commentary on EU General-Purpose AI Law

2. Substance

2.1. Article 93(1): Measures available to the Commission

5Article 93(1) AI Act sets out the specific measures that the Commission can, where ‘necessary and appropriate’, request from 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. These measures are listed as follows: request 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 ‘take appropriate measures to comply with the obligations set out in Articles 53 and 54’; to ‘implement mitigation measures’; and finally, ‘to restrict the making available on the market Article 3(10) AI Act: ‘making available on the market’ means the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge. , withdraw or recall the model’.

6The following analysis starts with a preliminary assessment of the legal character and formal conditions of the requests by the Commission to adopt measures under Article 93(1) (Section 2.1.1.) before analysing the substance of Article 93(1) (Sections 2.1.2. to 2.1.4.).

2.1.1. Article 93(1): Character and conditions of requests by the Commission

2.1.1.1. Informal versus formal requests

7Article 93 provides that the Commission ‘may request’ 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 adopt measures specified in Article 93(1)(a) to (c), without specifying the form that such requests will take. The only reference to a specific legal act can be found in Article 93(3), which indicates that the Commission ‘may by decision’ make binding the commitments offered 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. during the structured dialogue to implement mitigation measures. Arguably, the wording in Article 93 of ‘request’ in English, ‘demander’ in French, or ‘chiedere’ in Italian is not particularly strong or authoritative. Nonetheless, it has been argued by other authors that requests under Article 93(1) constitute decisions within the meaning of Article 288 TFEU,¹⁸ which are legally binding on their addressee. It follows that, as decisions, these requests can be subject to an action for annulment under Article 263 TFEU by the GPAI 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. .¹⁹ This is a convincing proposition, since the failure to comply with a measure requested by the Commission under Article 93 may ultimately lead to a fine under Article 101.²⁰ Recital 164 AI Act furthermore states that ‘[c]ompliance with the obligations should be enforceable, inter alia, through requests to take appropriate measures’,²¹ supporting the conclusion that such requests must be considered to have binding force. It should be noted that it is settled case law that the content of an act, not its nature or form, is relevant to the question of whether it is legally binding.²² Accordingly, the Court of Justice of the European Union (CJEU) will carefully examine the substance of an act and assess its effects to determine whether it produces binding legal effects.²³ Even emails or letters can qualify as decisions.²⁴ 

8At the time of writing (November 2025), the Commission has announced in its Guidelines on the scope of the obligations 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 that ‘[m]ore detailed acts will follow to further specify how these powers [laid down in Articles 91, 92, 93 and 101] will be implemented, in particular the implementing acts under Articles 92(6) and 101(6) AI Act.’²⁵ Article 92(6) provides that the Commission shall adopt implementing acts setting out the ‘detailed arrangements and the conditions for the evaluations’ under Article 92, while Article 101(6) refers to implementing acts containing ‘detailed arrangements and procedural safeguards for proceedings in view of the possible adoption of decisions’ under Article 101. This reference to implementing acts in these provisions is intriguing. Pursuant to Article 291(2) TFEU, the Commission may be empowered to adopt implementing acts where ‘uniform conditions for implementing legally binding Union acts are needed’. As discussed elsewhere in this commentary,²⁶ the rationale for implementing acts, as delineated in the case law of the CJEU, is to ensure that a legislative act is implemented under uniform conditions in all Member States.²⁷ However, as indicated above, the obligations 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. under Chapter V AI Act are enforced not by Member States but exclusively by the Commission. The reference to implementing acts in the context of enforcement by the Commission is not unprecedented. The Digital Markets Act (“DMA”), adopted before the AI Act, also sets out the possibility for the Commission to adopt implementing acts in situations that appear to go beyond their traditional use.²⁸ Whilst a more comprehensive analysis is beyond the scope of the present contribution, it is nevertheless worth noting that these references to implementing acts in a context where the enforcement lies with the Commission could pave the way for a new rationale for implementing acts.²⁹ 

9The requests under Article 93(1) which constitute a decision, and which can therefore be contested by an action for annulment, must be distinguished from requests formulated during structured dialogues. Pursuant to Article 93(2), 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 initiate a structured dialogue with 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. before the adoption of a formal request in the form of a decision by the Commission. As part of this structured dialogue, 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. could issue requests to 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. , for example to gather more information. Article 92(7) provides a clearer illustration of this by stating that 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 initiate a structured dialogue ‘to gather more information on the internal testing of the model, internal safeguards for preventing 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. , and other internal procedures and measures 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 taken to mitigate 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. ’. In the context of Article 93(2), 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. could discuss and invite 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 adopt specific measures to ensure compliance with its relevant obligations under the AI Act. It is suggested that these requests would not, prima facie, constitute decisions within the meaning of Article 288 TFEU, as interpreted by the CJEU. The Commission has indicated in the current Guidelines on the scope of obligations of GPAI models, under the section dedicated to supervision, investigation, enforcement, and monitoring, that 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. ‘will take a collaborative, staged, and proportionate approach’.³⁰ The possibility to initiate structured dialogues by 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. , as foreseen throughout Articles 91 to 93, appears to represent an important element of this collaborative approach.³¹ 

10If it is evident that structured dialogues must be distinguished from binding requests under Article 93(1), it remains unclear whether and in what ways they differ.³² Article 101 does not specifically provide for fines for failing to respond to an invitation by 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. to initiate a structured dialogue, which could indicate that structured dialogues are not legally binding. Article 101(1)(c) refers generally to the failure to comply with measures requested under Article 93 as a whole.³³ This could be interpreted as including requests made during structured dialogues. However, this interpretation is based on the assumption that the term ‘measure requested’ is not limited to those adopted under Article 93(1) by a Commission decision. This appears unlikely considering that Article 93(2) states that 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 initiate a structured dialogue ‘[b]efore a measure is requested’. This suggests that ‘measures requested’ should be interpreted as the measures requested by the Commission under Article 93(1). Moreover, from the perspective of legal certainty a fortiori in the context of fines, this interpretation is not entirely satisfactory, given that Article 101 does not specifically refer to structured dialogues. As the legislator provides no clarification on the concept of structured dialogues, one possible interpretive approach is to draw an analogy with the Commission’s simple requests for information issued in the context of competition law³⁴ or the Digital Services Act (“DSA”).³⁵ Under these instruments, simple requests represent a possible alternative for the Commission instead of issuing requests by decision, and 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. are not obliged to answer simple requests.³⁶ Reasoning by analogy to these simple requests, requests made as part of structured dialogues would not be binding, in the sense that 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 not be fined for not answering as this is not explicitly provided for in Article 101 AI Act. This would further be consistent with the reference in the current Guidelines to the Commission’s ‘collaborative, staged, and proportionate approach’.³⁷

11It is foreseeable that less formal, non-binding exchanges could take place outside of structured dialogues (given that structured dialogues are, at least prima facie, optional). We refer to such non-binding exchanges herein as ‘informal exchanges’. Informal exchanges are not mentioned in the AI Act. However, the practice of informal exchanges can be readily observed in competition law.³⁸ In light of the emphasis of the current Guidelines on the collaborative and staged approach, informal exchanges could play an important role in practice in the Commission’s enforcement strategy. At the time of writing in November 2025, an argument for this finds its ground in the Commission’s current Guidelines which state that 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. ‘encourages close informal cooperation with 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. during the training of their 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. to facilitate compliance and ensure timely market placement, in particular 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. 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. ’.³⁹ The same segment of the Guidelines indicates that 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. ‘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 collaborate and actively engage with 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.  at the later stages of potential formal proceedings conducted in line with 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. ’s powers under the AI Act (for example, through structured dialogues)’.⁴⁰ It is apparent from that sentence that structured dialogues form part of formal proceedings. By contrast, informal exchanges could facilitate a more flexible means of cooperation between 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 a GPAI 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. . In and of themselves, informal exchanges would not have any legal effect but could promote the spirit of cooperation that 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. seeks to engender with 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. . To the extent that there is 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. of obfuscation of the distinction between informal exchanges and structured dialogues, structured dialogues should be given a clear and well-defined purpose to ensure that they do not become redundant. In particular, this contribution argues below that, in contrast to informal exchanges, a structured dialogue demands a heightened degree of formality and that a ‘soft’ duty to cooperate arises as part of the structured dialogues.⁴¹

2.1.1.2. Formal requirements of requests

12Article 93 does not specify any formal requirements to be satisfied by requests to adopt measures. By contrast, Articles 91(4) and 92(7) AI Act both respectively indicate that requests for information or to conduct an evaluation must state the legal basis and purpose of the request, and, in the case of a request for information, specify the required information. They must also set out the period within which the request must be implemented and the fines applicable under Article 101 for non-compliance.⁴² Even in the absence of any express requirements in Article 93, generally, the Commission remains bound by the general principles of EU law and the Charter of Fundamental Rights of the European Union (“CFR”), and more specifically, pursuant to Article 94 AI Act, 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. enjoy the procedural rights conferred on the economic operators Article 3(8) AI Act: ‘operator’ means a provider, product manufacturer, deployer, authorised representative, importer or distributor. under the MSR.⁴³ These latter requirements include the right to be given reasons, to be informed of remedies and their respective time limits, and to be heard.⁴⁴ They reflect the general principles recognised in the case law of the CJEU as well as the right to good administration protected by the CFR. According to Article 18 MSR, these requirements apply to ‘[a]ny measure, decision or order’ which would exclude non-binding requests made as part of structured dialogues by, or informal exchanges with, the Commission at any time, as explained above.

13As part of its duty to provide reasons, the Commission must disclose clearly and unequivocally the reasons leading to the adoption of the measure.⁴⁵ This requirement is fundamental not only for the addressee to understand the decision, but also to control its legality.⁴⁶ This supposes, in the context of the measures adopted under Article 93, that the Commission specifies the infringement of the obligation under the AI Act and justifies the remedial action to be taken sufficiently to explain what is required specifically 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. . The duty to state the reasons includes, under the case law of the CJEU, the obligation to indicate the legal basis of a measure.⁴⁷ In a case concerning a request for information in the context of competition law, the Court has annulled a decision that did not satisfy the duty to state reasons. In particular, the Court noted that the decision contested did ‘not disclose, clearly and unequivocally, the suspicions of infringement which justify the adoption of that decision and does not make it possible to determine whether the requested information is necessary for the purposes of the investigation.’⁴⁸ 

14Despite the application of the procedural guarantees from the MSR and the existing body of case law from the CJEU, the procedural guarantees applicable to the exercise of Article 93 are less specific than those relating to Articles 91 and 92. For example, the reference to the requirement to state the possibility for fines provided for in Article 101 for failure to comply with the measure requested is absent in the context of Article 93. It has been argued by other commentators that such a requirement should apply by analogy based on a systemic interpretation of Article 93 in light of Articles 91(4) and 92(7).⁴⁹ Such an interpretation favours consistency between Articles 91 to 93 and legal clarity for 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. , especially as a differentiation in the procedural safeguards between those articles seems difficult to justify in principle.

15As to the right to be heard, it is settled case law that any person whose interests are affected by a decision by a public authority is to be afforded the opportunity effectively to make their view known.⁵⁰ The right is expressly included in the procedural safeguards of the MSR which, by virtue of Article 94 AI Act, apply mutatis mutandis to GPAI 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. subject to the exercise of Article 93 AI Act. Specifically, Article 18(3) MSR requires that 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 given the opportunity to be heard ‘within an appropriate period of not less than 10 working days’ unless the urgency of the measure requires an immediate action.⁵¹ The urgency of the measure should be ‘based on health or safety requirements or other grounds relating to the public interests covered by the relevant Union harmonisation legislation.’⁵² In such 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. shall be given the possibility to be heard as soon as possible thereafter and the measure shall be reviewed promptly by the Commission.⁵³ 

16Pursuant to Article 52 CFR, any restriction to the right to be heard must be provided for by law, not undermine the essence of that right, and should be proportionate. Thus, the Commission may only derogate from the right to be heard if the objective of the measure adopted under Article 93(1) cannot be achieved through a less, as suitable, restrictive measure. Neither the MSR nor the AI Act provides guidance on when a measure may qualify as ‘urgent’, nor what ‘other grounds relating to the public interests’ could be. However, the notion of urgency is well developed in the context of interim measures in the jurisprudence of the CJEU. Under that case law, interim measures may only be adopted where they are urgent to avoid ‘serious and irreparable damage to the party seeking their adoption, or which is intolerable for the public interest.’⁵⁴ The General Court has clarified that the condition of urgency is not an autonomous condition but ‘one aspect of the condition concerning 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 serious and irreparable damage.’⁵⁵ Thus, the notion of urgency must be analysed in relation to the need for an interim measure to prevent serious and irreparable damage.⁵⁶ According to the case law, it is not necessary for the occurrence of the damage to be demonstrated with absolute certainty to establish the existence of serious and irreparable damage.⁵⁷ Rather, it is sufficient that ‘the damage be foreseeable with a sufficient degree of probability’.⁵⁸ While this case law was developed in the specific context of the adoption of interim measures, by analogy, it may guide the assessment needed under Article 18 MSR applied mutatis mutandis in relation to Article 93 AI Act. Following that reasoning by analogy, the Commission could 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 take a measure under Article 93(1) without first giving the opportunity to be heard to 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. where such a measure is urgent to prevent reasonably foreseeable serious and irreparable damage to a public interest covered by the AI Act. This includes health, safety, and fundamental rights enshrined in the CFR, including democracy, the rule of law, and environmental protection.⁵⁹ This scenario could arguably be triggered by a GPAI model presenting 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. that was not mitigated at all, or not sufficiently mitigated to an acceptable level, in order to avoid serious and irreparable damage.

2.1.1.3. Article 93 and the principle of proportionality

17Article 93(1) conditions the exercise of the measures by the Commission to where it is ‘necessary and appropriate’. This explicit reference serves as a reminder that the Commission is bound by proportionality as a general principle when exercising its powers.⁶⁰ In accordance with this principle, Commission measures must be suitable and necessary to achieve the legitimate objective pursued by the legislation at issue, and the burden imposed should not be disproportionate in relation to the desired goal.⁶¹ A breach of the principle of proportionality can constitute a ground to challenge a decision before the Court.⁶²

18The principle of proportionality will be central in the determination of the use and scope of the measures available to the Commission. First, as foreshadowed above, the measure must be suitable to achieve the legitimate aim pursued by the AI Act. The relevant legitimate aim is the effective enforcement of 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, provided for under Articles 91 to 93 AI Act because the powers conferred under Articles 91 to 93 are intended to allow the Commission to ensure compliance with the obligations imposed on 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 under the AI Act. In turn, this contributes to the broader objective of the AI Act, which is to ensure ‘a high level of protection of health, safety, fundamental rights enshrined in the Charter, including democracy, the rule of law and environmental protection, against the harmful effects of AI systems 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. in the Union’, laid down in Article 1(1) AI Act. Secondly, a measure is necessary where there are no less restrictive and equally effective alternatives to achieve the aim pursued.⁶³ Where there is a choice between several appropriate measures, the Commission must choose the measure that restricts least the affected right, and, where relevant, the least onerous measure.⁶⁴ This question is relevant not only when determining the actions available among Articles 91, 92 and 93, but also within the different measures available to the Commission under Article 93(1)(a) to (c). It follows, for example, that the Commission may not request 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 recall the model under Article 93(1)(c) should a request to 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 adopt ‘appropriate measures’ under Article 93(1)(a) be equally effective to achieve the desired objective. Finally, the measure must be proportionate stricto sensu; that is, the disadvantages caused must not be disproportionate to the aims pursued by the measure.⁶⁵ In light of the case law of the CJEU, it follows that the measure imposed should not lead to a disproportionate burden for the GPAI 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. .⁶⁶ However, the mere fact that a measure involves a ‘significant workload’ 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. will not be sufficient by itself to establish that it is disproportionate.⁶⁷ The principle of proportionality in this context guarantees that the exercise by the Commission of its enforcement powers will not unduly compromise the rights 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. , including the freedom to conduct a business protected under Article 16 CFR.

2.1.2. Article 93(1)(a): Request to adopt appropriate measures

19Turning to the content of Article 93, Article 93(1)(a) sets out the possibility for the Commission to ‘request 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 take appropriate measures to comply with the obligations set out in Articles 53 and 54’. The reference to ‘appropriate measures’ again draws on the proportionality principle, particularly its first limb, which requires that a measure be suitable for attaining the desired outcome.

20It has been suggested by other authors⁶⁸ that ‘appropriate measures’ should be understood as ‘corrective action’, defined in Article 3(16) MSR as ‘any action taken by an economic operator Article 3(8) AI Act: ‘operator’ means a provider, product manufacturer, deployer, authorised representative, importer or distributor. to bring any non-compliance to an end where required by a market surveillance authority Article 3(26) AI Act: ‘market surveillance authority’ means the national authority carrying out the activities and taking the measures pursuant to Regulation (EU) 2019/1020. or on the economic operator Article 3(8) AI Act: ‘operator’ means a provider, product manufacturer, deployer, authorised representative, importer or distributor. ’s own initiative’. Such an interpretation makes sense, considering that the drafting of the AI Act draws on many aspects of the MSR; for example, Article 75(1) is particularly noteworthy.⁶⁹ In the specific context of Section 5 of Chapter IX, drawing from the MSR is most evident in Article 94 AI Act, which refers to Article 18 MSR applying ‘mutatis mutandis’. This close relationship between the AI Act and the MSR can be explained by the fact that the AI Act, at least in some respects,⁷⁰ follows the New Legislative Framework (“NLF”),⁷¹ which is a package of legislative measures, including the MSR, intended to increase product safety and improve market surveillance.⁷² However, returning to the meaning of ‘appropriate measures’, it is contended that the MSR’s definition of ‘corrective action’ includes not only the measures under Article 93(1)(a), but rather extends to all measures outlined in Article 93(1). Indeed, as demonstrated in Article 16 MSR, ‘corrective action’ requested by market surveillance authorities Article 3(26) AI Act: ‘market surveillance authority’ means the national authority carrying out the activities and taking the measures pursuant to Regulation (EU) 2019/1020. (“MSAs”) includes measures such as ‘preventing the product from being made available on the market Article 3(10) AI Act: ‘making available on the market’ means the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge. ’⁷³ and ‘withdrawing or recalling the product immediately and alerting the public to 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. presented’.⁷⁴ These measures are similar to those outlined in Article 93(1)(c), which allow the Commission to request 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 ‘restrict the making available on the market Article 3(10) AI Act: ‘making available on the market’ means the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge. , withdraw or recall the model’. It follows that all measures in Article 93(1) arguably fall under the definition of ‘corrective measures’ within the meaning of Article 3(16) MSR.

21Article 93(1) empowers the Commission to require specific actions from the GPAI 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. to ensure compliance with the AI Act.⁷⁵ Under Article 93(1)(a), these requests are limited to the obligations set out in Articles 53 and 54, which concern certain 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 and the authorised representatives Article 3(5) AI Act: ‘authorised representative’ means a natural or legal person located or established in the Union who has received and accepted a written mandate from a provider of an AI system or a general-purpose AI model to, respectively, perform and carry out on its behalf the obligations and procedures established by this Regulation. 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, respectively. In particular, Article 93(1)(a) does not refer to Article 52 which concerns the procedure for the classification of a GPAI model 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. and, perhaps even more surprisingly, to Article 55 regarding the obligations 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. . Based on the wording of Article 93(1)(a), it follows that the Commission cannot request 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 adopt appropriate measures in relation to their obligations under Articles 52 or 55. The consequences of restricting Article 93(1)(a)’s scope to Articles 53 and 54 have been debated by other commentators.⁷⁶ One interpretation argues that in light of the wording of the provision, the Commission may only rely on the more restrictive measures listed under Article 93(1)(c) in relation to other infringements under the AI Act,⁷⁷ such as infringements of Article 55. Other authors have argued that Article 93(1)(a) should apply, by analogy, to Article 55 as well.⁷⁸ This second interpretation underscores the necessity to ensure effective enforcement and highlights that 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. are more dangerous than models that are only subject to the obligations laid down in Article 53.⁷⁹ In addition, not all infringements of Article 55 would be covered by Article 93(1)(b) and/or (c), which are subject to specific conditions, a point examined further below.⁸⁰ The interpretation by analogy to extend Article 93(1)(a) to cover Article 55 would not be precluded, according to the authors, by the prohibition against punishment by analogy in criminal law (as derived from the maxim nulla poena sine lege) as it would neither modify the existence of criminal liability nor its conditions.⁸¹ An additional argument in favour of this proposition could be derived from the observation that Article 93(1)(a) offers the least intrusive of the measures outlined in Article 93(1) and it would be attractive, including from the perspective of proportionality, for the Commission to have the wider range of responses to breaches of Article 55 that Article 93(1)(a) offers. However, it is doubtful whether an interpretation that includes Article 55 within Article 93(1)(a) would survive judicial review by the CJEU. Indeed, the principle of legality, which is a corollary of the principle of legal certainty and is recognised as a general principle in EU law,⁸² requires ‘the law to give a clear definition of offences and the penalties which they attract’.⁸³ It is settled case law that this principle does not only apply to criminal law provisions but also to ‘specific administrative instruments imposing or permitting the imposition of administrative penalties, and that it applies not only to provisions establishing the elements which constitute an offence, but also to those specifying the consequences arising from an offence’.⁸⁴ Should 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. be fined under Article 101(1)(c) for a failure to comply with a measure requested under Article 93(1)(a) which is applied by analogy to Article 55, that fine could potentially be challenged on the basis of lack of legal certainty. In conclusion, the interpretation by analogy to extend Article 93(1)(a) to Article 55 infringements should be rejected.

22There are several reasons that may explain the absence of reference to Articles 52 and 55 in Article 93(1)(a). Firstly, the absence of a reference to Article 52 is accounted for by the fact that this provision outlines the procedure for classifying a GPAI model 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. and 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 notify the Commission.⁸⁵ It is unclear how Article 93(1)(a) could be ‘activated’ in relation to non-compliance with Article 52, or what could constitute a relevant ‘appropriate measure’ in this context, considering that the Commission can request information and documentation under Article 91, which covers ‘any additional information necessary for assessing 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 compliance with this Regulation’. Furthermore, it is important to note that the Commission under the AI Act has the ability to designate a model 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. .⁸⁶ This could be considered a more appropriate response to a failure 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. to notify in comparison to the adoption of a request under Article 93(1)(a).

23The absence of a reference to Article 55 in Article 93(1)(a) is more curious. One may tentatively consider that a violation of Article 55 is considered so serious that the Commission should proceed directly in such cases to use the powers indicated in Article 93(1)(c) (i.e. restrict the making available to the market, withdraw or recall the model). However, this argument is not compelling. The Article 93(1)(c) measures should be regarded as ultima ratio in view of the degree of their interference and in accordance with the principle of proportionality.⁸⁷ Under the principle of proportionality, the Commission may only rely on the more restrictive measures set out in Article 93(1)(c) if the objective pursued cannot be appropriately achieved through the powers available to the Commission under Article 93(1)(a)–(b). Another possible explanation is that the obligations set out in Article 55 are sufficiently covered in Article 93(1)(b). This provision, which is analysed below, allows the Commission to request that 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. ‘implement mitigation measures, where the 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’.⁸⁸ As indicated above, it has been argued by authors that not all infringements of Article 55 are covered by Articles 93(1)(b) and (c) as they are subject to specific conditions.⁸⁹ Thus, this explanation is not entirely satisfactory either.

24The heart of the matter is whether an enforcement gap presents itself in relation to Article 55, and, if so, whether there are any means to close that gap in some way. Analysing Article 55 closely, the provision contains the following 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. : to perform model evaluation; to assess and mitigate possible 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. ; to keep track of, document and report information about 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 possible corrective measures; and to ensure an adequate level of cybersecurity protection.⁹⁰ Starting with the obligation in Article 55(1)(a) and (b), it can be noted that under Article 93(1)(b) the Commission can request 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 ‘implement mitigation measures’ where 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. performed an evaluation of the model under Article 92 and concluded that there is ‘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. ’. Article 93(1)(b) thus allows the Commission to act 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. did not correctly evaluate, assess and mitigate possible 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. , thus failing to comply with its obligations under Articles 55(1)(a) and (b).⁹¹ As regards the third obligation, that is to keep track of, document, and report relevant information about serious accidents and possible measures to correct them (Article 55(1)(c)), 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 be subject to a request for documentation and information pursuant to Article 91. As regards in particular the obligation to keep track of serious accidents and possible measures to correct them, it could be argued that should the Commission request such information 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. fails to comply with that request, the Commission could issue a fine for failing to comply with that request under Article 101(1)(b) (or in any case issue a fine under Article 101(1)(a)). This would indirectly enforce the obligation to keep track of the relevant information under Article 55(1)(c). Finally, regarding the fourth obligation (Article 55(1)(d)), it can be argued that the request to adopt ‘mitigation measures’ could also cover the obligation to ensure an adequate level of cybersecurity protection which is closely linked to mitigation measures 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. .⁹² The link between cybersecurity and 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. appears more clearly in the GPAI Code of Practice which states, as part of Commitment 6 of the Safety and Security Chapter in the GPAI Code of Practice, that: ‘Signatories commit to implementing an adequate level of cybersecurity protection for their models and their physical infrastructure along the entire model lifecycle, as specified in the Measures for this Commitment, to ensure 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. stemming from their models that could arise from unauthorised releases, unauthorised access, and/or model theft are acceptable (pursuant to Commitment 4).’⁹³ Seen in this light, there would be no enforcement gap with regard to Article 55.

25Another issue raised in other commentaries is whether the Commission can request preventive measures under Article 93(1)(a) to stop an imminent infringement.⁹⁴ It has been argued by other authors that the wording of this provision is sufficiently broad to include preventive measures and that the possibility of requesting preventive measures would be consistent with the risk-based approach of the EU AI Act and its objective of ensuring a high level of protection per Article 1(1) AI Act.⁹⁵ The same authors also emphasise that such measures must be proportionate.⁹⁶ Considering the scope of Article 93(1)(a), this possibility would only apply to the prevention of imminent infringements of Articles 53 and/or 54. However, it is unclear what measure could be requested that would prevent an infringement of Articles 53 and/or 54. If the Commission requires 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 adopt ‘appropriate measures to comply with the obligations set out in Articles 53 and 54’, this arguably implies that 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. had failed to comply with these obligations, for example by neglecting to draw up and maintain the technical documentation of the GPAI model.⁹⁷ However, it is difficult to see how this could be sufficiently anticipated by the Commission for it to recognise that a preventive measure should be requested to stop the infringement from occurring.⁹⁸ Moreover, this could call into question the characterisation of ‘appropriate measures’ that anticipate an infringement as ‘corrective action’ under the MSR. Indeed, as observed above, the MSR defines corrective action as ‘any action taken by an economic operator Article 3(8) AI Act: ‘operator’ means a provider, product manufacturer, deployer, authorised representative, importer or distributor. to bring any non-compliance to an end’ where required by a market surveillance authority Article 3(26) AI Act: ‘market surveillance authority’ means the national authority carrying out the activities and taking the measures pursuant to Regulation (EU) 2019/1020. or on the economic operator Article 3(8) AI Act: ‘operator’ means a provider, product manufacturer, deployer, authorised representative, importer or distributor. ’s own initiative, which suggests that there is a priori a situation of non-compliance.

2.1.3. Article 93(1)(b): Request to adopt mitigation measures

26Article 93(1)(b) confers the power on the Commission to request that 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. ‘implement mitigation measures, where the 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’. Under Article 92, 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. or independent experts appointed by the Commission can perform an evaluation of the GPAI model to assess compliance of 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. with its obligations under the AI Act or 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. . Article 93(1)(b) is the logical consequence of an evaluation that has identified unacceptable 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 have not, or not sufficiently, been mitigated. At the outset, the wording in the English version of Article 93(1)(b), ‘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. ’, could create the misconception that the ‘serious and substantiated concern’ relates to the existence 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. . As noted by other commentators, the existence 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. ⁹⁹ is a necessary condition but is not sufficient to justify a request to adopt mitigation measures.¹⁰⁰ This risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. must raise a serious and substantiated concern. The formulation in French and German more clearly conveys that the serious and substantive concern is not ‘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. but ‘regarding’ such risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. .¹⁰¹ Such an interpretation is more consistent with the fact that under the AI Act 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 identify, assess and mitigate those risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. , but they are not required to eliminate them entirely.¹⁰² This is clearly reflected in the Safety and Security Chapter of the GPAI Code of Practice, for example, which gives shape to the obligations under Article 55. Under Commitment 4 in that Chapter, signatories commit to determine whether 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. stemming from the GPAI model are acceptable and, in particular, to only proceed with ‘the development, the making available on the market Article 3(10) AI Act: ‘making available on the market’ means the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge. , and/or the use of the model, if 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. stemming from the model are determined to be acceptable’.¹⁰³ Moreover, requiring 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 eliminate 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. entirely would go against the logic of the AI Act. Indeed, should 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 classified as a 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. pursuant to Article 51 be required to eliminate 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. entirely, a distinct category 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. would no longer be necessary.

27As already noted, the concern related to 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. identified must be ‘serious and substantiated’. The words ‘serious and substantiated concern’ are not defined in the AI Act. As regards the term ‘serious concern’, the wording suggests at first sight a certain level of gravity. Other authors have suggested that a serious concern exists where ‘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. parameters of the severity of the damage as well as the probability of a damage occurring, either individually or in combination, are to be assessed as completely unacceptable because the situation thereby created is absolutely intolerable in light of the high level of protection that must be guaranteed.’¹⁰⁴ It is argued that, contrary to the authors’ proposition, the assessment of whether a serious concern exists must necessarily take into account both the probability and severity of the damage in question, in order to mirror the definition of ‘ risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. ’ in the AI Act: ‘the combination of the probability of an occurrence of harm and the severity of that harm’.¹⁰⁵ To consider only the severity of the damage occurring would result in an unduly broad interpretation. For example, biological, radiological and nuclear risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. or loss of control¹⁰⁶ would already reach the threshold of ‘serious concern’ by only considering the severity of the harm. A possible analogy can be drawn from the definition of a ‘product presenting a serious risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. ’ found in the MSR. The latter defines it as a product ‘for which, based on 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. assessment and taking into account the normal and foreseeable use of the product, the combination of the probability of occurrence of a hazard causing harm and the degree of severity of the harm is considered to require rapid intervention by the market surveillance authorities Article 3(26) AI Act: ‘market surveillance authority’ means the national authority carrying out the activities and taking the measures pursuant to Regulation (EU) 2019/1020. , including cases where the effects of 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. are not immediate’.¹⁰⁷ It is noteworthy that the definition includes situations where the effects of 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. are not immediate. To ensure legal certainty, further guidance from the Commission or the CJEU would be needed to determine the meaning of ‘serious concern’.

28As to the term ‘substantiated’, it supposes that the concern is based on evidence. The term ‘substantiated concern’ can be found in other areas of secondary EU law. For example, it is defined in the EU Deforestation Regulation as a ‘duly reasoned claim based on objective and verifiable information regarding non-compliance with this Regulation and which could require the intervention of competent authorities’.¹⁰⁸ Similarly, Regulation 2024/3015 on the prohibition of products derived from forced labour describes it as a ‘reasonable indication based on objective, factual and verifiable information for the Commission or competent authorities to suspect that it is likely that a product was made with forced labour’.¹⁰⁹ Interpreted mutatis mutandis, there is a compelling argument that a ‘substantiated concern’ in the context of Article 93(1)(b) AI Act could mean a duly reasoned claim based on objective, factual and verifiable information regarding 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.

29Finally, as regards the scope of the mitigation measure, the latter should not go beyond what is necessary to achieve the objective pursued under the principle of proportionality, explored in more detail above.¹¹⁰ It follows in particular that the Commission can only require 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. to be reduced to a level that is acceptable, but not to eliminate it altogether.¹¹¹ Should there be a choice between several appropriate measures to reduce this risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. to an acceptable level, the Commission must adopt the least restrictive one. It follows that a measure which would go beyond reducing that risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. to an acceptable level will not be proportionate. The determinative question here is what can be considered an acceptable level and how this can be measured, as this is not defined in the AI Act. The Safety and Security Chapter of the GPAI Code of Practice also leaves 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. discretion in that respect. Specifically, under Measure 4.1, 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. are required to describe and justify their method to consider 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. as acceptable, including by defining appropriate 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. and/or other criteria as part of their Safety and Security Framework.¹¹² In time, the limits of this discretion may be established by further Commission guidance or a decision of the CJEU.

2.1.4. Article 93(1)(c): Request to restrict, withdraw or recall model

30Under Article 93(1)(c), the Commission may request 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 restrict the making available on the market Article 3(10) AI Act: ‘making available on the market’ means the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge. , withdraw or recall the model’. These measures echo the powers conferred on market surveillance authorities Article 3(26) AI Act: ‘market surveillance authority’ means the national authority carrying out the activities and taking the measures pursuant to Regulation (EU) 2019/1020. under the MSR.¹¹³ Both the position in Article 93(1) as the last measure in this provision and the character of the interference indicate that a restriction, withdrawal or recall represents an ultima ratio tool for the Commission to enforce the AI Act.¹¹⁴ As highlighted above, respecting the principle of proportionality will be of particular importance in the exercise of these powers by the Commission. In particular, the objective to be pursued should be achieved by the least restrictive measure available to the Commission, in accordance with the principle of proportionality.¹¹⁵ For example, this could be done by requesting 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 adopt appropriate measures or mitigation measures under Article 93(1)(a) and (b). It is foreseeable that the Commission could rely on Article 93(1)(c), for example, 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. fails to take earlier, appropriate corrective action or where the non-compliance or 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. persists despite 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. having already engaged with the relevant GPAI 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. about the issue.¹¹⁶ The principle of proportionality also plays a role in the selection of the particular measure from Article 93(1)(c)’s three options; the three measures set out therein have different levels of impact, with the request to recall the model causing the most interference, as explained further below.¹¹⁷

31As regards the possibility for the Commission to request that 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. ‘restrict the making available on the market Article 3(10) AI Act: ‘making available on the market’ means the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge. ’, the absence of a definition of such a measure leaves a certain margin of discretion to the Commission, subject to the principle of proportionality. The term ‘ making available on the market Article 3(10) AI Act: ‘making available on the market’ means the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge. ’ is defined in Article 3(10) as ‘the supply of 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. or 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. for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge’. The word ‘restrict’ indicates that the Commission may impose certain limitations on the GPAI 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. . At the time of writing (November 2025), before the Commission’s powers in Article 93 have become applicable, it is difficult to predict which specific measures the Commission could request to achieve the relevant restriction, especially considering that an AI model is not a tangible product. This will depend on the circumstances of the case. It has been argued by some commentators that this provision is more likely to cover ‘functional’ restrictions on the model, concerning its use, application or performance, although the wording does not exclude other kinds of restrictions such as geographical, temporal or quantitative.¹¹⁸ Temporal restrictions could be a useful tool, for example in the context of elections, where restricting the making available of an GPAI model by imposing certain conditions could be a valuable way of mitigating 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. to democratic processes.¹¹⁹ Restrictions could also entail technical measures, for example to require 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 include certain filters into their GPAI model to restrict certain outputs.¹²⁰ It seems premature to suggest that one type of restriction will be more likely or prevalent than another as the measure will need to respond specifically, and therefore proportionately, to the infringement in question, on a case-by-case basis.

32As to the withdrawal of the model under Article 93(1)(c), its meaning can be derived mutatis mutandis from the AI Act’s definition of ‘ withdrawal of an AI system Article 3(17) AI Act: ‘withdrawal of an AI system’ means any measure aiming to prevent an AI system in the supply chain being made available on the market. ’ as ‘any measure aiming to prevent 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. in the supply chain being made available on the market Article 3(10) AI Act: ‘making available on the market’ means the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge. ’.¹²¹ As for the recall of the model, likewise, Article 3(16) AI Act defines ‘ recall of an AI system Article 3(16) AI Act: ‘recall of an AI system’ means any measure aiming to achieve the return to the provider or taking out of service or disabling the use of an AI system made available to deployers. ’ as any measure aiming ‘to achieve the return to 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. or taking out of service or disabling the use of 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. made available to deployers Article 3(4) AI Act: ‘deployer’ means a natural or legal person, public authority, agency or other body using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity. ’¹²². Both definitions are strongly inspired by the MSR, in accordance with the New Legislative Framework.¹²³ The definitions quoted indicate that these measures cover different situations. In the context of a withdrawal, the model has not yet been made available to the downstream provider Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. .¹²⁴ This could also cover a situation where the model is not made available to a downstream provider Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. but is integrated 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. 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. .¹²⁵ A request to withdraw targets the distribution of the model and can, for example, entail an interdiction on selling the model.¹²⁶ In contrast, in the context of a recall, the GPAI model will have already been made available to the downstream provider Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. . 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 then take the necessary action ‘to achieve the return’ of the model, which, in the context of a model that is not a tangible product, would mean removing the ability to use it. For example, this could entail that 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. revokes the licence, disables the model through technical means or adopts other measures to take the model out of service. 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. , recalling a model could be particularly resource-intensive. Precisely how burdensome and costly this will be will depend on the circumstances.

2.2. Article 93(2): Structured dialogue

33Pursuant to Article 93(2), 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 initiate structured dialogues with 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. before the adoption of a measure under Article 93(1).¹²⁷ The AI Act provides little indication on the notion and structure of structured dialogues.¹²⁸ The only outcome of a structured dialogue expressly provided for in the AI Act figures in Article 93(3), examined below.¹²⁹ This provision allows the Commission to declare that commitments offered by 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 adopt mitigation measures to address 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. are binding.

34The possibility to initiate a structured dialogue before the Commission adopts coercive measures under Article 93(1) is similar to other informal mechanisms that precede formal enforcement proceedings.¹³⁰ As noted above, this echoes in particular the enforcement structure in competition law¹³¹ and the DSA,¹³² where the Commission can issue so-called ‘simple requests’ before adopting a binding request via decision. To date, the Commission Guidelines offer little further guidance on the content of structured dialogues under the AI Act. It has been argued by other commentators that structured dialogues are voluntary.¹³³ According to this view, 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. can neither be compelled to respond to a structured dialogue nor are subject to an obligation to cooperate.¹³⁴ Indeed, the AI Act does not express specific obligations arising during structured dialogues, and in particular does not set out fines should 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. refuse to respond to an invitation by 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. to engage in a structured dialogue.¹³⁵ This could suggest that structured dialogues are voluntary and cannot give rise to 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. . Consequently, following this view, should 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. not cooperate, 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. could close the dialogue and issue a binding decision. However, another interpretation is possible in light of the duty of cooperation set out in Article 53(3) AI Act, which provides that 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 ‘cooperate as necessary with 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. in the exercise of their competences and powers pursuant to this Regulation’. Therefore, it could be contended that, in the absence of express language to this effect in the AI Act, even if 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. cannot be compelled to respond to the initiation of a structured dialogue by 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. , in light of Article 53(3), if they choose to engage with a structured dialogue, they must cooperate. It could be argued that this requires 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. to act in good faith during structured dialogues and (at least) not to provide false or misleading information 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. .¹³⁶ This interpretation has the advantage of enhancing the utility and effectiveness of structured dialogues and minimises the opportunity 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 exploit structured dialogues for the purpose of prolonging non-compliance.

2.3. Article 93(3): commitments by 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.

35A structured dialogue initiated under Article 93(2) can come to a conclusion with different outcomes: the structured dialogue can be brought to an end without further action; the Commission can decide to adopt a binding decision under Article 93(1); or the Commission can, by decision, make certain commitments binding under Article 93(3).¹³⁷ In accordance with Article 93(3), if during the structured dialogue initiated under Article 93(2), 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 a GPAI model 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.  ‘offers commitments to implement mitigation measures to address 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, the Commission may, by decision, make those commitments binding and declare that there are no further grounds for action’ (henceforth referred to as ‘commitment decision’). The possibility of offering commitments that can be made binding through a commitment decision is only included in Article 93 and only in relation to mitigation measures.  

36Article 93(3) gives structured dialogues a practical effect by allowing the Commission to reach a negotiated outcome to its enforcement efforts when 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. ’ commitments adequately address its concerns. Thus, the commitment decision represents a valuable instrument to encourage and ensure compliance. The possibility for the Commission to adopt commitment decisions can be found in other pieces of secondary legislation such as Regulation 1/2003¹³⁸ and the DSA.¹³⁹ Indeed, the practice of adopting commitment decisions is particularly well developed in competition law,¹⁴⁰ which can be expected to inform the practice of the Commission under the AI Act given that the AI Act contains very few specifications regarding Article 93(3) commitments and no recital referring to commitment decisions that could inform their interpretation. As is apparent from their use in competition law, commitment decisions can offer a number of benefits, including enhanced procedural efficiency and streamlined implementation, given that the commitments are established by 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. themselves.¹⁴¹ One particular benefit 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. , seen in the competition law context,¹⁴² is that a commitment neither represents an admission of guilt nor constitutes a formal finding of a violation under the AI Act.

37The discussion that follows explores the scope and conditions of commitments under the AI Act, drawing upon the practice of commitments in competition law and the case law of the CJEU therein, when relevant. It then turns to the scenario of the Commission rejecting commitments, as well as the review of commitments and the legal effects of commitment decisions. The analysis will draw attention to the distinctive features of commitment decisions under the AI Act, such as their limited scope and the lack of legal basis in the AI Act to issue a fine 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. ’s breach of a commitment decision.

2.3.1. Scope and conditions of commitments offered under Article 93(3)  

38Article 93(3) only allows commitments offered by 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 be made binding in relation to mitigation measures to address 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 the level of the Union. This supposes a scenario in which concerns have been raised during the structured dialogue about 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. not being sufficiently mitigated 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. . It is unclear why the legislator did not extend the possibility of using commitment decisions to ensure compliance with other relevant provisions of this Regulation. By comparison, Article 71(1) DSA is much broader and provides the possibility to offer ‘commitments to ensure compliance with the relevant provisions of this Regulation’. Hypothetically, one could envisage, for example, a commitment to adopt appropriate measures to put in place (an adequate) policy to comply with Union law on copyright and related rights in pursuant to Article 53(1)(c) AI Act. It has been suggested by other authors that 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 also offer commitments with regard to other measures and, in such cases, the Commission may informally (and outside the scope of Article 93(3)) determine not to impose binding measures by decision or fines if these commitments are correctly implemented.¹⁴³ The suggestion is fortified by the fact that, in the area of competition law, the Commission continues to settle cases informally, outside of the commitment decision procedure codified in Regulation 1/2003.¹⁴⁴ However, such informal settlements do not provide the Commission or 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. with legal certainty as they lack the clear legal binding effect of a decision. 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. could potentially derive an argument from the principle of legitimate expectations, under the condition that they can demonstrate that the Commission has provided them with clear assurances which have led them to entertain justified expectations.¹⁴⁵ It remains to be seen whether the Commission will in its AI Act enforcement practice allow and accept commitments outside the case of mitigation measures to address 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 non-formal basis.

39It is also noteworthy that, unlike a request to adopt mitigation measures under Article 93(1)(b), Article 93(3) does not refer to an evaluation carried out under Article 92 which ‘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’. Consequently, the possibility of agreeing to a commitment does not depend on the Commission demonstrating that there is a ‘serious and substantiated concern’ regarding 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. .

40Regarding the form, the AI Act does not specify the formal requirements for commitments made 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. . In the absence of specific requirements, the form of the commitments offered 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. under this provision is free, as long as its intention to offer commitments under Article 93(3) is clear. In competition law, commitments are usually monitored by appointing a third party approved by the Commission and referred to as the ‘Monitoring and/Divestiture Trustee’.¹⁴⁶ It is possible that a similar process for commitments under the AI Act will be followed, where an entity approved by the Commission could be appointed to supervise and evaluate 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 adherence to the commitments and to report its evaluation of 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 conduct to the Commission.

2.3.2. Rejection of commitments by the Commission

41In light of the wording of Article 93(3), the Commission may, but is not obliged to, accept the commitments offered 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. . The Commission may reject commitments if it considers that the commitments offered are inadequate to mitigate 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. .¹⁴⁷ This raises the question whether the Commission is required to provide reasons when rejecting commitments. In contrast to the DSA, the AI Act does not expressly require the Commission to reject commitments offered by 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. in a ‘reasoned decision’.¹⁴⁸ Under the case law of the CJEU, the obligation to state reasons applies, as a general rule, to all EU acts that ‘produce legal effects’.¹⁴⁹ Moreover, only measures that produce binding legal effects capable of bringing a distinct change in the legal position of applicants constitute challengeable acts.¹⁵⁰ Two cumulative arguments suggest that the rejection of a commitment does not constitute an act that produces legal effect. Firstly, since the possibility to offer commitments 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. is voluntary (and only becomes binding once accepted by the Commission), the rejection of the commitments offered 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. would only result in the Commission or 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. continuing its investigation. Secondly, and related to this first point, there is no right 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 have commitments accepted by the Commission. Thus, the legal position of 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. and in particular the rights to engage in work and to pursue an occupation and to conduct a business protected respectively by Articles 15 and 16 CFR are not affected. Therefore, it is unlikely that the Commission’s rejection 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 proposed commitment will be recognised as a legal act in the sense of the case law.  

42Nevertheless, this issue could arise if 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. were to challenge the Commission’s request to adopt measures under Article 93(1) or a fine imposed under Article 101(1). It is foreseeable that, as part of such a challenge, 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. might rely on the fact that, as part of the structured dialogues, it had offered commitments to mitigate 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. that were rejected. More specifically, when fixing a fine, the commitments offered 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. could be relied on as a mitigating factor. Such an argument might support a challenge to the proportionality of the measure or the fine under Article 263 TFEU. Indeed, it has been suggested by some authors that, in accordance with the principle of proportionality, the Commission may only 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 adopt measures under Article 93(1) if 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. cannot equally effectively be mitigated through a commitment decision under Article 93(3) following a structured dialogue under Article 93(2).¹⁵¹ As indicated above, the principle of proportionality requires EU institutions not to exceed the limits of what is necessary in order to attain the legitimate objectives pursued, and should there be a choice between several equally effective measures, recourse must be had to the least onerous.¹⁵² This does seem to suggest that a proportionality challenge in the scenario under consideration could have merit. However, reasoning by analogy with competition law, where the CJEU has exercised limited judicial review to grant a wide discretion to the Commission to make a proposed commitment binding or reject it,¹⁵³ indicates that such a challenge would face difficulties.

2.3.3. Control by the Commission

43Another topic concerns the control of the Commission of the commitments offered 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. . As indicated in the previous subsection, the Commission can reject commitments that are not adequate to mitigate 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. . In contrast, what if 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. commits beyond what is necessary? Is the Commission obligated to reject or recast the commitments offered in this scenario? The precise scope of the obligations of the Commission under the principle of proportionality when adopting commitment decisions had been highly contested in the area of competition law¹⁵⁴ but was largely settled in the Alrosa case.¹⁵⁵ In that decision, the CJEU confirmed that the Commission is bound by the principle of proportionality when adopting a decision making commitments offered by undertakings binding.¹⁵⁶ However, the Court indicated that observance of this principle in the context of Article 9 Regulation 1/2003, which concerns commitment decisions, is limited to verifying that ‘the commitments in question address the concerns it expressed to the undertakings concerned and that they have not offered less onerous commitments that also address those concerns adequately.’¹⁵⁷ In particular, it follows that the Commission is not required ‘itself to seek out less onerous or more moderate solutions than the commitments offered to it’.¹⁵⁸ Moreover, the Court in Alrosa limited the judicial review to manifest errors of assessment.¹⁵⁹ 

44Reasoning by analogy, it follows that, when assessing commitment decisions under the AI Act, the Commission need only consider whether the commitments are sufficient to mitigate 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 it will benefit from a generous margin of discretion to make that assessment. It is not required to identify less onerous or more moderate solutions than those offered by 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. . This reasoning takes stock of the fact that the commitments are formulated and offered by 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. , not required by the Commission. Moreover, the decision whether a commitment can be considered suitable to mitigate 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. arguably involves a highly technical assessment by the Commission. In that regard, it is settled case law that the Court grants the Commission discretion in ‘administrative procedure entailing complex technical evaluation’.¹⁶⁰

2.3.4. Legal effects of a commitment decision, including the potential for fines for non-compliance

45Turning to the legal effects of a Commission’s commitment decision under Article 93(3), it is clear from the language of the article that commitments offered by 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. become binding. The article also states that, in this scenario, the Commission is to ‘declare that there are no further grounds for action’ such that the commitment decision leads to the closure of the case by the Commission.¹⁶¹ It is suggested that the commitment decision creates legitimate expectations 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. that the Commission will not pursue its investigation or request measures such as those under Article 93(1)(c).¹⁶² Therefore, 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 perspective, a key incentive to offer commitments is the possibility of avoiding further Commission enquiries and/or the more restrictive measures under Article 93(1)(c), whereby the Commission can request that 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. restrict the model’s availability on the market or withdraw or recall it.

46A key question remains whether the phrase ‘no further grounds for action’ excludes the possibility of a fine under Article 101. In competition law, it is well recognised that commitments allow undertakings to avoid a possible fine¹⁶³ except if they fail to comply with their commitments.¹⁶⁴ The CJEU has declared that the adoption of a commitment decision ‘closes infringement proceedings against those undertakings by allowing them to avoid a finding of an infringement of competition law and a possible fine.’¹⁶⁵ This compounds the incentive 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 offer commitments. In the AI Act, the terms ‘no further grounds for action’ could be interpreted as excluding the issuing of a fine by the Commission under Article 101(1). However, Article 101(1) indicates that when issuing a fine, the Commission shall also take ‘into account commitments made in accordance with Article 93(3)’.¹⁶⁶ In light of the wording of Article 101, it seems that the Commission’s acceptance of a commitment does not exclude the possibility of a fine altogether, whilst the existence of a commitment will be taken into account for setting the level of a fine.¹⁶⁷ This could, in practice, have a discouraging effect on 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. ’ inclination to offer commitments. However, as Ambrock and Raji have suggested, the waiving of a fine could be made conditional upon the implementation of the commitment 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. .¹⁶⁸ 

47This possibility for the Commission to waive a fine under Article 101 so long as 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. implements correctly the commitments acquires particular importance because, somewhat surprisingly, the AI Act does not provide explicitly for the possibility to issue a fine if 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. fails to implement its commitments. Article 101(1)(a) does provide for the possibility to fine 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. for an infringement of the relevant provision of this Regulation, but it is highly unlikely that the Commission could rely on this article for a breach of a ‘binding’ commitment 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. without going against the requirement of legal certainty. As indicated, Article 101(1) states that commitments should be ‘taken into account’ when fixing the fine. This, for example, contrasts with the DSA¹⁶⁹ and Regulation 1/2003¹⁷⁰ which both explicitly provide for such possibility should the regulatee fail to comply with a commitment made binding by a decision. In such cases, the Commission is only required to demonstrate that the regulatee did not comply with the commitments, but is not required to establish a breach of the relevant provisions. It has been argued that the Commission could rely on Article 101(1)(c) pursuant to which the Commission may impose a fine if 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. fails to comply with a measure requested under Article 93.¹⁷¹ However, there are two arguments against using Article 101(1)(c) to justify a fine 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. ’s failure to comply with its binding commitment. Firstly, Article 101(1)(c) refers to ‘a measure requested under Article 93’. As commitments are offered by 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. rather than requested by the Commission, it is difficult to interpret a ‘measure requested’ as covering a commitment declared binding under Article 93(3). Moreover, a systematic reading of Article 93 goes against this interpretation. Indeed, Article 93(2) refers to structured dialogues ‘before a measure is requested’. Therefore, the commitments made by 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. during such dialogues arise before a measure is requested under Article 93(1).

48In any case, the Commission can fine 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. under Article 101(1)(a) for infringing the relevant provisions of the AI Act. The failure to implement commitments may be taken into account as an aggravating factor under Article 101(1). The main difference, compared with the DSA or Regulation 1/2003 powers to fine, is that, under Art 101(1) AI Act, the Commission will need to prove a breach of the AI Act. In light of this, making a fine waiver conditional upon the correct and timely implementation by 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 their commitments will be central to ensuring the effectiveness of commitments as a solution to promote compliance as well as legal certainty.

49Finally, it is worth considering the possibility for the Commission to take further action, including the possibility to review commitments. In contrast again to the DSA¹⁷² or Regulation 1/2003,¹⁷³ Article 93(3) does not specify that the Commission may take further action if there has been a material change in any of the facts on which the commitment decision was based, if 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 acted contrary to its commitments, or if the decision was based on incorrect or misleading information 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. . Article 93(3) only indicates that the Commission may adopt a commitment decision and ‘declare that there are no further grounds for action’. Ambrock and Raji have argued that a parallel should be drawn with Article 71(2) DSA and that the Commission may also take further action in similar scenarios under the AI Act, at the very least if 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 failed to comply with its commitment.¹⁷⁴ The contrary would result in the Commission being unable to take action, even in cases where the decision was incorrect or 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. breached its commitments. The authors observe that this cannot have been the intention of the legislator.¹⁷⁵ The argument has merit; it is unattractive to interpret the terms ‘no further grounds for action’ to mean that the Commission cannot reopen the case in the scenarios envisaged by Article 71(2) DSA and Article 9(2) Regulation 1/2003 even though there is no equivalent express provision in the AI Act. The argument is not contradicted by the principle of legitimate expectations. Indeed, applying existing case law,¹⁷⁶ a legitimate expectation would not arise in cases in which the Commission’s acceptance of a commitment was procured by means of 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 false or incomplete information. Similarly, 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 not invoke the principle of legitimate expectations in order to prevent the Commission changing an ‘existing situation’ (here, the absence of further action undertaken by the Commission following the commitment decision) in the exercise of its discretion.¹⁷⁷ Finally, it may also be in the interest 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. to review and modify commitments should there be a material change in the facts on which the decision was adopted. Considering the absence of express provision to that effect in the AI Act, this possibility should be clearly indicated in the commitment decision in order to promote legal clarity.¹⁷⁸ To err on the safe side, the Commission could adopt the decision under these conditions; that is, the conditions are set constitutively by the decision itself.