Article 51: Classification of GPAI Models with Systemic Risk

Article 51 answers the question: when is a GPAI model considered to have systemic risk? Two routes:

Route 1 — High-impact capabilities (Article 51(1)): A GPAI model is classified as systemic risk when it has high-impact capabilities evaluated using appropriate technical tools and methodologies, including indicators and benchmarks. The AI Office may designate on its own initiative or after a qualified alert from the scientific panel (Article 90).

Route 2 — Compute presumption (Article 51(2)): A GPAI model is presumed to have high-impact capabilities when the cumulative amount of computation used for its training, measured in floating-point operations (FLOPs), is greater than 10^25. The Commission may update this threshold by delegated act under Article 97 based on evolving technological benchmarks.

Providers can rebut the presumption with substantiated arguments through the Article 52 procedure—the threshold is not absolute, but the burden is on the provider.

Once a GPAI model is classified as systemic risk (by notification or designation):

1. Article 55 additional obligations activate: model evaluation, adversarial testing (red teaming), tracking and reporting serious incidents, ensuring adequate cybersecurity protection. 2. Annex XI Section 2 documentation becomes mandatory: detailed evaluation strategies, red-teaming measures, and system architecture. 3. The model appears on the AI Office's published list of systemic-risk GPAI models. 4. Providers may need to adhere to relevant codes of practice under Article 56 or demonstrate equivalent compliance.

The 10^25 FLOPs presumption was calibrated to frontier models at the time of legislative negotiations. It is a rebuttable presumption, not a hard classification boundary—models below the threshold can still be designated if they demonstrate high-impact capabilities through other evidence.

As training efficiency improves (distillation, architecture advances, data quality), the compute proxy may under- or over-capture risk. The Commission has delegated-act power to adjust the threshold and the Annex XIII criteria to reflect state of the art.

• Article 52 — Procedure for classification (notification, AI Office designation, rebuttal, publication).
• Article 53 — Baseline GPAI obligations (apply to all GPAI providers, regardless of systemic risk).
• Article 55 — Additional obligations activated by systemic-risk classification.
• Article 56 — Codes of practice as a compliance pathway for systemic-risk obligations.
• Annex XI — Section 2 documentation triggered by systemic-risk status.
• Annex XIII — Criteria and thresholds for systemic-risk classification (including the 10^25 FLOPs presumption).
• Article 90 — Scientific panel that may trigger qualified alerts for AI Office classification.
• Article 97 — Delegated acts for updating the Annex XIII threshold.
• Article 101 — Penalties for GPAI providers.
• Article 113 — Application dates.

1. A general-purpose AI model shall be classified as a general-purpose AI model with systemic risk if it meets any of the following conditions:

(a) it has high impact capabilities evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks;

(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.

2. A general-purpose AI model shall be presumed to have high impact capabilities pursuant to paragraph 1, point (a), when the cumulative amount of computation used for its training measured in floating point operations is greater than 10^25.

3. The Commission shall adopt delegated acts in accordance with Article 97 to amend the thresholds listed in paragraphs 1 and 2 of this Article, as well as to supplement benchmarks and indicators in light of evolving technological developments, such as algorithmic improvements or increased hardware efficiency, when necessary, for these thresholds to reflect the state of the art.

The recitals in the same consolidated AI Act on EUR-Lex contextualise systemic risk, high-impact capabilities, the 10^25 FLOPs calibration, the scientific panel role, and the Commission's delegated-act update mechanism. Use the official preamble on EUR-Lex—do not rely on unofficial recital lists without checking sequence and wording against the authentic text.