Skip to main content
Praxikon

How a general-purpose AI model is classified as a model with systemic risk: Article 52 of the AI Act

··25 min read

A Dutch municipality builds an assistant that helps staff answer questions about permits. Under the hood runs a language model from a European model developer, whose weights are publicly available and which the municipality runs on its own servers. That language model is a general-purpose AI model (Article 3, point (63)). One day the European Commission designates it as a model with systemic risk. Does anything change for the municipality?

How a model acquires that status is set out in Article 52, with the conditions in Article 51 and the criteria in Annex XIII. This piece goes through Article 52 paragraph by paragraph and then looks at it from the municipality's side: what changes in its own role, and what must the model developer now do and provide? A bank and an insurer follow, and briefly education. The examples are fictitious.

In brief

A model becomes a model with systemic risk in one of two ways (Article 51(1)). The first is automatic: the model has high impact capabilities, which is presumed above a training compute of 10^25 floating point operations (paragraph 2). The provider then notifies the Commission without delay and in any event within two weeks, and may argue, with sufficiently substantiated arguments, that exceptionally its model does not present systemic risks (Article 52(1) and (2)). If the Commission rejects those arguments, the model is considered a model with systemic risk (paragraph 3).

The second way is a decision of the Commission, ex officio or following a qualified alert from the scientific panel, on the basis of the criteria in Annex XIII (paragraph 4). A provider whose model is designated in that way may request reassessment at the earliest six months later (paragraph 5). The Commission ensures that a list of models with systemic risk is published and kept up to date (paragraph 6).

The Digital Omnibus on AI (Regulation (EU) 2026/1744) did not amend Article 51, Article 52 or Annex XIII. They are in Chapter V and have applied since 2 August 2025 (Article 113).

A designation does not change the role of a municipality or bank that builds the model into its own system. It does mean that the model developer has to do more (Article 55), and that the exceptions for models released under a free and open-source licence no longer apply (Article 53(2) and Article 54(6)).

What the Commission can then do to supervise such a model is covered in how the Commission supervises general-purpose AI models. The news on the start of enforcement is in GPAI enforcement has started.

The framework: Article 51 and Annex XIII

Systemic risk means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain (Article 3, point (65)). High-impact capabilities are capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models (Article 3, point (64)).

Article 51(1) sets two conditions; meeting one is enough:

  • point (a): the model has high impact capabilities evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks;
  • point (b): based on a decision of the Commission, ex officio or following a qualified alert from the scientific panel, it has capabilities or an impact equivalent to those set out in point (a), having regard to the criteria set out in Annex XIII.

A model is presumed to have high impact capabilities when the cumulative amount of computation used for its training, measured in floating point operations, is greater than 10^25 (Article 51(2)). The Commission adopts delegated acts, when necessary, to amend the thresholds and to supplement benchmarks and indicators so that they reflect the state of the art (paragraph 3). The consolidated text of 27 July 2026 still says 10^25.

Annex XIII lists seven criteria the Commission takes into account: the number of parameters (a), the quality or size of the data set, for example measured through tokens (b), the amount of computation used for training, or a combination of other variables such as estimated cost, time or energy consumption of the training (c), the input and output modalities (d), the benchmarks and evaluations of capabilities, including the level of autonomy and scalability and the tools the model has access to (e), a high impact on the internal market due to its reach, which is presumed when it has been made available to at least 10 000 registered business users established in the Union (f), and the number of registered end-users (g). According to recital 111, the decision is taken on the basis of an overall assessment of those criteria.

The routes at a glance

RouteWho starts itWhat the text requiresWhen the obligations apply (Commission guidelines)
Notification after the threshold (paragraph 1)ProviderNotify within two weeks, with the information necessary to demonstrate that the requirement has been metFrom the moment the model meets Article 51(1), point (a)
Counter-arguments (paragraphs 2 and 3)Provider, with the notificationSufficiently substantiated arguments that exceptionally the model does not present systemic risksRejected: from the moment of the threshold. Accepted: the obligations cease once the provider is informed
Designation without notification (paragraph 1, third sentence)CommissionA model presenting systemic risks of which the Commission has not been notifiedFrom the moment the model met Article 51(1), point (a) (this is how the Commission reads this route, paragraph 43)
Designation based on Annex XIII (paragraph 4)Commission, ex officio or following an alert from the panelEquivalent capabilities or impact, assessed against Annex XIIIFrom the moment the provider is informed of the decision
Reassessment (paragraph 5)Provider of a designated modelReasoned request, with objective, detailed and new reasons, at the earliest after six monthsNot addressed

The last column is not in Article 52 itself but in the Commission guidelines on the scope of the obligations for providers of general-purpose AI models, paragraphs 26, 40 and 44. We use the formally adopted version of 19 November 2025 (C(2025) 7719 final). The text whose content the Commission approved on 18 July 2025 has different paragraph numbers in a number of places. The guidelines are not binding; they show how the Commission applies the provision.

Paragraph 1: notification within two weeks

Where a model meets the condition referred to in Article 51(1), point (a), the provider notifies the Commission without delay and in any event within two weeks after that requirement is met or it becomes known that it will be met. The notification includes the information necessary to demonstrate that the requirement has been met (Article 52(1)).

In our reading, the second trigger, "it becomes known that it will be met", matters most. Training takes considerable planning, including the upfront allocation of compute resources, so a provider can know whether its model will meet the threshold before training is completed (recital 112). The Commission infers from this that a notification may be required before training is complete: the provider estimates the training compute before starting the large pre-training run, notifies if the estimate meets the threshold, and otherwise monitors actual and expected compute usage over the course of training and throughout the model's lifecycle (guidelines, paragraphs 29 and 30). According to the guidelines, the notification covers in particular the estimated compute in floating point operations with two significant figures, and how it was estimated (paragraph 31).

The notification duty is tied only to the condition in point (a). In the absence of a delegated act, the Commission sees two situations that trigger it: the model meets or will meet the 10^25 threshold, or it results from a modification of a model with systemic risk in which the modification uses more than a third of the original training compute (guidelines, footnote 8 to paragraph 31, and paragraphs 60 and 67).

The third sentence of paragraph 1 closes the gap: if the Commission becomes aware of a model presenting systemic risks of which it has not been notified, it may decide to designate it as a model with systemic risk. The legal text speaks broadly of a model presenting systemic risks; in the guidelines the Commission limits this route to a model that met the condition of Article 51(1), point (a), and was not notified (paragraph 43). According to the guidelines, the obligations then apply from the moment the model met the condition, and the provider may be fined under Article 101 for failing to notify (paragraphs 43 and 44). That Article has applied since 2 August 2026.

Paragraphs 2 and 3: the provider's counter-arguments

With its notification, the provider may present sufficiently substantiated arguments to demonstrate that, exceptionally, although the model meets the requirement, it does not present, due to its specific characteristics, systemic risks and therefore should not be classified as a model with systemic risk (paragraph 2). The text says "with its notification": this is not a separate objection afterwards.

Where the Commission concludes that the arguments are not sufficiently substantiated and the provider was not able to demonstrate that the model does not present, due to its specific characteristics, systemic risks, it rejects those arguments and the model is considered a model with systemic risk (paragraph 3). Paragraph 3 does not say what happens if the Commission accepts the arguments.

The guidelines fill that in (paragraphs 32 to 40):

  • The burden is on the provider. Its arguments must be sufficiently substantiated and manifestly call the presumption into question (paragraph 36). It may argue that its model does not have high-impact capabilities, with benchmark results, including forecasted ones, and further information such as the architecture, the number of parameters, the training examples, the modalities and the expected tool use (paragraphs 33 and 34).
  • A moving bar. Which models are the most advanced shifts over time; the Commission does not consider the notion to refer to a fixed level of capabilities (paragraph 37).
  • Mitigation is not an argument. That a model does not present systemic risks because of mitigations already implemented or planned is, for the Commission, not a suitable ground for excluding it from classification; those mitigations belong to the risk mitigation under Article 55(1) (paragraph 39).
  • A hearing before rejection. Before a rejection, the provider receives a draft decision letter in advance and has access to its file; the decision states reasons (paragraph 35).
  • Arguments suspend nothing. On rejection, the obligations apply from the moment the model met the condition. On acceptance, they cease from the moment the provider is informed of the decision, but the provider must renotify the Commission if there is a substantial change in the facts or if the decision was based on substantially incomplete, incorrect or misleading information; an acceptance does not preclude a later designation under paragraph 4 (paragraph 40).

Paragraph 4: designation by the Commission

The Commission may designate a model as presenting systemic risks, ex officio or following a qualified alert from the scientific panel pursuant to Article 90(1), point (a), on the basis of the criteria set out in Annex XIII (paragraph 4, first subparagraph). A model below the 10^25 threshold can therefore also become a model with systemic risk, for example because of its reach or the tools it has access to. The Commission is empowered to amend Annex XIII by delegated act, specifying and updating its criteria (paragraph 4, second subparagraph; Article 97).

A qualified alert under Article 90(1), point (a), concerns a concrete identifiable risk at Union level. Article 90(1), point (b), also provides for an alert that a model meets the conditions referred to in Article 51, but Article 52(4) refers only to point (a). In our reading this makes little difference in practice: following an alert under point (b), the Commission can designate ex officio. What an alert is and how the panel works is covered in our piece on Articles 90 to 94. A municipality or bank cannot issue a qualified alert itself; only the panel can.

According to the guidelines, after a designation under paragraph 4 the obligations apply from the moment the provider is informed of the decision (paragraph 44). Article 52 does not provide for a prior hearing before a designation under paragraph 4; the guidelines mention the right to be heard expressly only for the rejection of arguments, and base it on Article 41 of the Charter of Fundamental Rights (paragraph 35). In our reading that right to be heard also applies to a designation under paragraph 4, because it covers any individual measure that adversely affects a person.

Paragraph 5: reassessment upon request

Upon a reasoned request of a provider whose model has been designated under paragraph 4, the Commission takes the request into account and may decide to reassess whether the model can still be considered to present systemic risks on the basis of the criteria in Annex XIII (paragraph 5). The request must contain objective, detailed and new reasons that have arisen since the designation decision. It can be made at the earliest six months after the designation decision; if the Commission maintains the designation, again at the earliest six months after that decision.

Two things stand out. The text says "may decide": the Commission must take the request into account but is not obliged to reassess. And paragraph 5 applies only to a designation under paragraph 4, not to the rejection of arguments under paragraph 3 or to a designation under paragraph 1. In our reading, in those cases, as with paragraph 4, an action for annulment before the Union courts within two months remains available (Article 263 TFEU), since such a decision is addressed to the provider and has legal effects. A provider whose arguments were accepted may still have to renotify (guidelines, paragraph 40).

Paragraph 6: the published list

The Commission ensures that a list of general-purpose AI models with systemic risk is published and keeps that list up to date, without prejudice to the need to observe and protect intellectual property rights and confidential business information or trade secrets (paragraph 6). The text does not say where or in what form, nor how quickly a designation must appear on the list. In our reading, the list covers both models that are models with systemic risk via the threshold and designated models, since paragraph 6 refers to general-purpose AI models with systemic risk without distinguishing by route.

Whether and where that list had been published as of 19 November 2026 is something we could not establish from the Commission sources we consulted. According to the Commission, models with systemic risk are currently developed by a handful of companies (General-Purpose AI Models in the AI Act: Questions and Answers). If you want to know whether a model is on the list, also ask its provider directly.

When does this apply?

  • The AI Act entered into force on 1 August 2024 (Article 113, first paragraph). Entry into force is not the same as application.
  • Articles 51 and 52 are in Chapter V and have applied since 2 August 2025, as has Chapter VII with the scientific panel (Article 113, third paragraph, point (b)). According to the guidelines, the Commission can take decisions under Article 52 from that date (paragraph 104).
  • Fines under Article 101 have been possible since 2 August 2026: Article 101 is excepted from 2 August 2025 and follows the general date of application (Article 113, second paragraph and third paragraph, point (b)).
  • Models placed on the market before 2 August 2025 must comply with the obligations by 2 August 2027 (Article 111(3)). How that transitional period affects the notification duty in Article 52(1) is not addressed separately in the text.
  • The Digital Omnibus on AI entered into force on 27 July 2026. It did not amend Articles 51 to 55 or Annex XIII.

Main example: the municipality with an assistant for permit questions

The municipality develops the assistant and puts it into service under its own name. It is the provider of an AI system (Article 3, point (3)) and a downstream provider (Article 3, point (68)). The model developer is the provider of the model.

Its own role does not change. A designation is addressed to the provider of the model. The municipality does not become a provider of a model with systemic risk, and its own obligations stay the same. It continues to ensure that staff interacting with the assistant are informed that they are interacting with an AI system, unless that is obvious (Article 50(1)), and continues its measures for AI literacy (Article 4). An assistant that helps staff answer questions is in principle not a high-risk system. If the municipality starts using it to assess whether residents are eligible for a benefit, that can change (Annex III, point 5(a)), with obligations from 2 December 2027, including registration of the use (Article 49(3); Article 26(8)) and a fundamental rights impact assessment (Article 27(1)). Because it builds the system itself and puts it into service under its own name, it is then also the provider of that high-risk AI system (Article 3(3)), with the obligations of Article 16, including the conformity assessment (Article 43) and registration of the system (Article 49(1)). For a system put into service before 2 December 2027, the rules apply only after a significant change in its design, and for public authorities by 2 August 2030 at the latest (Article 111(2)). In the Netherlands, a market surveillance authority for Annex III has not yet been designated.

What the model developer must now do. In addition to the obligations in Articles 53 and 54, it performs model evaluation including adversarial testing, assesses and mitigates possible systemic risks at Union level, keeps track of, documents and reports without undue delay to the AI Office and, as appropriate, to national competent authorities relevant information about serious incidents and possible corrective measures, and ensures an adequate level of cybersecurity protection for the model and its physical infrastructure (Article 55(1)). It may rely on a code of practice within the meaning of Article 56 (Article 55(2)), such as the General-Purpose AI Code of Practice, whose Safety and Security chapter the Commission describes as relevant only to providers subject to the obligations of Article 55.

What it must provide. This is where the tangible difference lies for this municipality. A model released under a free and open-source licence with publicly available parameters is normally exempt from the obligation to draw up technical documentation and to provide information to downstream providers, but that exception does not apply to general-purpose AI models with systemic risk (Article 53(2)). The model developer must therefore draw up, keep up to date and make available information and documentation to providers that integrate the model into their systems, such as the municipality, without prejudice to the protection of trade secrets. That information enables them to have a good understanding of the capabilities and limitations of the model and to comply with its own obligations, containing at a minimum the elements set out in Annex XII (Article 53(1), point (b)). Had it been a developer established in a third country, it would now also need an authorised representative in the Union (Article 54(1) and (6)). If the municipality uses a closed model through an API, the information duty already applied; in our reading, the designation should then be reflected in updated information, since Article 53(1), point (b), requires that information to be kept up to date.

What it contributes itself. The reach criterion in Annex XIII, point (f), counts registered business users established in the Union. In our reading, a municipality only counts if it is registered with the model developer as a business user, for example through an API account; one that merely downloads the public weights and runs them itself usually is not. It cannot request a designation itself. If the model developer does not meet its obligations towards it, the municipality, as a downstream provider, can lodge a complaint (Article 89(2)).

When it becomes a model provider itself. If it trains the model further. According to the Commission, that only happens if the modification leads to a significant change in the model's generality, capabilities or systemic risk, with as an indicative criterion more than a third of the original training compute (guidelines, paragraphs 59 and 60). For a model with systemic risk, the modified model is then presumed to have high-impact capabilities, and the municipality must notify the Commission itself under Article 52(1) (paragraphs 67 and 68). In our reading, a municipality that only adapts the model to its own texts does not reach that point.

The privacy side. The assistant processes questions containing personal data. The usual obligations continue to apply, such as a DPIA where the processing is likely to result in a high risk (Article 35 GDPR). In our reading, what helps most there is the information the model developer makes available under Article 53(1), point (b), and Annex XII of the AI Act. What it documents on systemic risks it reports to the AI Office (Article 55(1)); there is no obligation to share that with the municipality. Recital 110 does mention, among other things, harming privacy and harmful bias as systemic risks.

Example 2: the bank and the insurer

A bank builds an assistant, on the model of a large provider accessed through an API, that helps staff summarise complaint letters. If that model is designated, or turns out to be a model with systemic risk via the threshold, the bank's own role does not change either. The model provider must, however, assess and mitigate systemic risks and report serious incidents to the AI Office (Article 55(1)). In our reading, the bank should record the model's status in its outsourcing and continuity policy, and agree by contract that the provider informs it of a notification, designation or reassessment.

If the bank starts using the system to evaluate the creditworthiness of natural persons, it is in principle a high-risk system (Annex III, point 5(b)), with obligations from 2 December 2027, including a fundamental rights impact assessment (Article 27(1)). A bank that builds and uses the system itself is then both provider and deployer of that system (Article 3(3) and (4)). The market surveillance authority is then in principle the financial supervisory authority, in so far as the use is directly connected with the provision of the financial services (Article 74(6)). The same applies, via Annex III, point 5(c), to an insurer that builds a model into risk assessment and pricing in relation to natural persons in the case of life and health insurance. Whether the underlying model has systemic risk does not change that classification of the system.

Briefly: education

A university of applied sciences that builds a study assistant on a model with systemic risk is in the same position as the municipality: its role stays the same, and the model developer bears the obligations of Article 55. If it uses the assistant to evaluate learning outcomes, it may be high-risk (Annex III, point 3(b)), with obligations from 2 December 2027. Emotion recognition in education is prohibited, except for medical or safety reasons (Article 5(1)(f)), whatever model runs underneath.

What to record now

  1. Which model, from whom. Per application: model name and version, the provider and its authorised representative in the Union, and whether the model is released under a free and open-source licence.
  2. Systemic risk or not. Ask the provider whether it has notified under Article 52(1), whether its model has been designated and whether it is on the list under Article 52(6). Record the date of the answer.
  3. The documentation. What you received under Article 53(1), point (b), and Annex XII, and whether it was updated after a notification or designation.
  4. Contract. An agreement that the provider informs you of a notification, the outcome of counter-arguments, a designation or a reassessment.
  5. Your own modifications. Whether you train the model further, with how much compute, and whether that brings you close to the criterion in the guidelines.

Want to know which AI Act and GDPR obligations come into play for your application and what role your organisation has? With the case check you describe the application in a few questions.

Frequently asked questions on the designation of AI models with systemic risk

Sources

EUR-Lex: Regulation (EU) 2024/1689 (AI Act), original text, recitals 110, 111 and 112 (12 July 2024, accessed 19 November 2026)
EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI); Articles 51 to 55 and Annex XIII not amended (8 July 2026, published 24 July 2026, accessed 19 November 2026)
European Commission: General-Purpose AI Models in the AI Act: Questions and Answers (accessed 19 November 2026)
European Commission: The General-Purpose AI Code of Practice (10 July 2025, accessed 19 November 2026)
EUR-Lex: Treaty on the Functioning of the European Union, consolidated version, Article 263 (OJ C 202, 7 June 2016, accessed 19 November 2026)
EUR-Lex: Regulation (EU) 2016/679 (GDPR), Article 35 (27 April 2016, accessed 19 November 2026)

Newsletter

Every Tuesday, the AI Act week ahead in 5 minutes

A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.

Practical and short · No spam · One-click unsubscribe

Written by

Zahed Ashkara, freelance AI and privacy consultant and legal adviser

Zahed Ashkara

Freelance AI and privacy consultant and legal adviser · AI governance & safety

Zahed Ashkara is a freelance AI and privacy consultant and legal adviser for the GDPR and the EU AI Act. Organisations hire him as AI governance consultant, privacy counsel, external DPO or AI governance lead, have him carry out a DPIA, FRIA or AI Act classification, or book an in-company training. He focuses on AI governance and safety, including AI agents. He works exclusively as a freelancer, through Embed AI (trade name Praxikon). He is a Certified AI Compliance Officer (CAICO) and has been a member of the NEN standards committee Artificial Intelligence & Big Data since July 2026.

Follow on LinkedInView expert profile

Use these sources in your AI assistant too: Praxikon in Claude and other assistants

Real-World Stories

Discover how organizations apply this in practice

Related articles

How the Commission supervises general-purpose AI models: Articles 90 to 94 of the AI Act

Articles 90 to 94 of the AI Act set out four steps towards providers of general-purpose AI models: a qualified alert from the scientific panel to the AI Office, followed by the Commission's request for documentation and information, an evaluation of the model and a request for measures, with procedural rights for the provider. What each paragraph says, how Articles 101 and 113 come in, and what a municipality, bank or university of applied sciences that builds such a model into its own system will notice and record.

EU AI ActArticle 90Article 91

Codes of conduct under Article 95 of the AI Act: voluntary rules for AI that is not high-risk

Article 95 of the AI Act lets organisations draw up codes of conduct on a voluntary basis: to apply the requirements for high-risk systems to AI that is not high-risk, and to set their own objectives on ethics, the environment, AI literacy, inclusive design and vulnerable groups. What such a code does and does not do, who may draw it up, how it differs from the codes of practice under Articles 50 and 56 and from codes of conduct under the GDPR. With a partnership of municipalities as the main example.

EU AI ActArticle 95Codes of conduct