A general-purpose AI model is classified as presenting systemic risk where it has high-impact capabilities, assessed under Article 51. The Act creates a presumption of systemic risk where cumulative training compute exceeds 10 to the power of 25 floating point operations, and it also allows the Commission to designate a model on the basis of the criteria in Annex XIII. The provider can seek to rebut the presumption with sufficient evidence, and classification can be reassessed as capabilities and evidence change.
What does systemic risk mean in Article 51?
Systemic risk in this context is risk at Union level that arises from the scale, reach and capabilities of a general-purpose AI model. The concern is not a single faulty output in one deployment, but effects that can propagate across many downstream systems, sectors and affected groups because the same model is widely integrated. Article 51 therefore looks at high-impact capabilities and their potential to create large-scale harm, including through misuse, loss of control, or widespread deployment in consequential contexts. That framing is broader than a single system's accuracy score. A provider assessing classification should describe the model's capabilities, its foreseeable reach, and the kinds of downstream use that could amplify an effect, and should record why those factors do or do not point to systemic risk on the facts of that model.
How does the training compute presumption work?
Article 51 provides that a general-purpose AI model is presumed to present systemic risk where the cumulative amount of computation used for its training, measured in floating point operations, is greater than 10 to the power of 25. The presumption is rebuttable: a provider may present sufficiently substantiated arguments that, despite meeting the compute figure, the model does not present systemic risk because of its specific characteristics. The compute figure is therefore a starting point for classification, not a final legal conclusion on its own. Providers should keep a defensible record of how training compute was estimated, what was included in the estimate, who checked it, and what other capability evidence was considered alongside it. A figure without a method, a scope note and a reviewer is difficult to rely on if the classification is questioned.
For the broader regulatory framework, see General-purpose AI models overview.
What other designation routes exist?
The Commission may also designate a general-purpose AI model as presenting systemic risk on its own initiative or following a qualified alert, using the criteria in Annex XIII. Those criteria look beyond a single compute number to matters such as the number of users, the scale and reach of use, the model's capabilities and impact, and other indicators of high-impact capability. Designation can therefore apply to a model below the compute presumption threshold where the Annex XIII criteria are met. Providers should not treat falling below the compute figure as a safe conclusion without considering the designation route. The record should note whether any Annex XIII indicator was considered, what evidence was available, and whether any alert or designation process is relevant to the model.
What happens under Article 52 after classification is in view?
Article 52 sets the procedure that follows a classification trigger. A provider that meets or expects to meet the Article 51 criteria must notify the Commission, and the Commission may designate a model, reassess a designation, and maintain information about designated models. The provider has opportunities within that procedure to present evidence and to request reassessment where circumstances change. The practical consequence is that classification is a procedural as well as a substantive question: timing, notification, evidence and follow-up all matter. A provider should diarise the notification step, keep the submission and any response together, and record any change in capability, deployment scale or evidence that could support a reassessment request. Delay or silence is not a neutral course where a notification duty has been triggered.
How should a provider document a classification decision?
A useful classification file names the model and version, states the compute estimate and method, summarises capability evidence, addresses each relevant Annex XIII consideration, and records the conclusion and the reviewer. It should also record the date of the assessment, the sources checked, and the trigger for the next review, such as a retraining run, a capability jump, a large increase in deployment, or new Commission guidance. Where the presumption is relied on or rebutted, the file should show the reasoning in full rather than stating a bare result. That file supports the Article 53 documentation, informs downstream providers where relevant, and gives a later reviewer a clear trail. It does not by itself prove that the classification is correct, and it should be updated when the underlying facts change.
Continue with Chapter V map for general-purpose AI models.
What does this guide not decide?
This guide does not decide whether any named or unnamed model presents systemic risk, what its training compute figure is, or whether a rebuttal would succeed. It does not interpret Annex XIII for a specific model, assess a notification, or predict a Commission designation. Those matters depend on model-specific evidence, the consolidated legal text, and the procedure in Articles 51 and 52 as applied to that model. A provider should confirm its classification position with a competent authority or legal adviser, preserve the evidence relied on, and treat classification as a live question to be revisited when capabilities, scale or guidance change.
Frequently asked questions
Does exceeding the compute threshold automatically mean systemic risk?
The threshold creates a presumption of systemic risk, not an irrebuttable conclusion. A provider may rebut it with sufficiently substantiated evidence relating to the specific characteristics of the model. The record should show both the compute estimate and the capability evidence relied on, and the classification should be confirmed through the Article 52 procedure where applicable.
Can a model below the compute threshold be designated?
Yes, in outline. The Commission may designate a model as presenting systemic risk on the basis of the Annex XIII criteria, including indicators of high-impact capability and scale, even where the compute presumption is not met. Providers should consider designation risk alongside the compute figure and keep the assessment under review.
Is classification a one-time decision?
No. Classification should be revisited when the model is retrained, its capabilities change materially, its deployment scale grows, or new guidance or designation criteria become relevant. A dated classification file with a stated review trigger helps a provider show that the question was kept live rather than treated as closed at first release.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗Commission AI Act Service Desk timeline ↗European Commission AI Act overview ↗