What is the systemic-risk threshold for GPAI models?

Cumulative compute used for training greater than 10 to the power of 25 floating-point operations creates a presumption of systemic risk. The Commission may update the threshold and may designate models below it based on capabilities and impact.

What extra obligations come with systemic risk?

Article 55 adds model evaluation with adversarial testing, assessment and mitigation of Union-level systemic risks, serious-incident tracking and reporting to the AI Office, and an adequate level of cybersecurity protection for the model and its infrastructure.

Can a provider contest a systemic-risk designation?

Yes. A provider whose model meets the compute presumption can argue that, because of its specific characteristics, the model does not present systemic risk, and the Commission assesses that reasoning. The obligations apply unless and until the argument succeeds.

EU AI Act glossary · Core concepts· Art. 51, 55

Systemic risk (GPAI)

A general-purpose AI model is presumed to carry systemic risk when the cumulative compute used to train it exceeds 10²⁵ floating-point operations, triggering stricter obligations under Article 55.

Systemic-risk models must undergo standardised model evaluation including adversarial testing, assess and mitigate risks at Union level, track and report serious incidents to the AI Office, and ensure an adequate level of cybersecurity for the model and its physical infrastructure. The 10²⁵ FLOP figure is a presumption, not a definition: the Commission can designate a model as carrying systemic risk on capability grounds even below the threshold, and a provider can argue that its model exceeds the threshold without presenting equivalent capability.

Where Systemic risk (GPAI) sits in the Regulation

Art. 51
Art. 52
Art. 55

Who it binds

Systemic risk (GPAI): why it matters

This is the only place in the Act where a quantitative training-compute figure creates legal consequences, and it changes the compliance posture from documentation to active safety engineering: red-teaming, model evaluation and incident reporting are ongoing operational commitments, not a file you assemble once. There is also a notification duty — a provider that knows its model will cross the threshold has to tell the Commission without delay, which means the trigger is felt during training rather than at release.

Worked example

Systemic risk (GPAI) in practice

A lab planning a frontier training run estimates cumulative compute during capacity planning and finds it will pass 10²⁵ FLOP. From that point it has to notify the Commission, stand up an adversarial-testing programme before release, and build an incident-reporting channel to the AI Office — work that has to start months ahead of the launch date, not at the point the model ships.

Correction

Systemic risk (GPAI): the mistake people make

Reading the threshold as the whole test. It is a rebuttable presumption in both directions: the Commission can designate a below-threshold model on the basis of capability, reach or impact, so a provider that stops measuring once it is under 10²⁵ FLOP may still be designated.

Questions

Systemic risk (GPAI): common questions

What is the systemic-risk threshold for GPAI models?: Cumulative compute used for training greater than 10 to the power of 25 floating-point operations creates a presumption of systemic risk. The Commission may update the threshold and may designate models below it based on capabilities and impact.
What extra obligations come with systemic risk?: Article 55 adds model evaluation with adversarial testing, assessment and mitigation of Union-level systemic risks, serious-incident tracking and reporting to the AI Office, and an adequate level of cybersecurity protection for the model and its infrastructure.
Can a provider contest a systemic-risk designation?: Yes. A provider whose model meets the compute presumption can argue that, because of its specific characteristics, the model does not present systemic risk, and the Commission assesses that reasoning. The obligations apply unless and until the argument succeeds.

Keep reading

Systemic risk (GPAI): terms that come up alongside it

General-purpose AI (GPAI) modelArt. 3(63)

The foundation-model category with its own obligation set under Articles 53 to 55.

GPAI model providerArt. 53–55

The model-layer role supervised by the AI Office rather than national authorities.

European AI Office

The Commission body that supervises GPAI models and shapes the Act's guidance.

Serious incident reportingArt. 73

The mandatory notification duty when a high-risk system causes serious harm.

See how Systemic risk (GPAI) applies to your systems

Run the free 2-minute EU AI Act checker — no sign-up.

Run the free check