EU AI Act Article 55 — GPAI Models with Systemic Risk

What Article 55 actually requires

Article 55 of Regulation (EU) 2024/1689 sets additional obligations on providers of GPAI models classified as carrying systemic risk under Article 51.

A GPAI model has systemic risk when it has high-impact capabilities evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks. The Regulation sets a default presumption of systemic risk when the cumulative amount of computation used for training, measured in floating-point operations, is greater than 10^25 FLOPs (Article 51(2)).

The Commission may designate models as systemic-risk based on capabilities, the number of parameters, the quality or size of the dataset, the input/output modality, the degree of autonomy, the impact on the internal market, the number of registered end users, etc.

The four additional obligations

Providers of systemic-risk GPAI models must, in addition to Article 53 duties:

• (a) Perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing of the model with a view to identifying and mitigating systemic risks.
• (b) Assess and mitigate possible systemic risks at Union level, including their sources, that may stem from the development, the placing on the market, or the use of GPAI models with systemic risk.
• (c) Keep track of, document, and report, without undue delay, to the AI Office and, as appropriate, to national competent authorities, relevant information about serious incidents and possible corrective measures to address them.
• (d) Ensure an adequate level of cybersecurity protection for the GPAI model with systemic risk and the physical infrastructure of the model.

Who is covered

Initially, frontier-lab providers (OpenAI, Anthropic, Google DeepMind, Meta, Mistral, etc.) whose state-of-the-art models exceed the 10^25 FLOP presumption threshold or are otherwise designated by the Commission.

What to do

• If your training run exceeds 10^25 FLOPs, notify the Commission without delay (Article 52(1)).
• Build a model-evaluation programme aligned with state-of-the-art protocols. The AI Office is publishing reference protocols.
• Run adversarial-testing campaigns on the model itself (red-teaming) and document outcomes.
• Maintain a serious-incident register tied to the Article 73 / 79 reporting flow.
• Cybersecurity-harden the model: weights protection, supply-chain integrity, physical security of the training infrastructure.

The AI Office is developing codes of practice under Article 56 to operationalise Article 55. Adherence to an approved code of practice creates a presumption of compliance.

Inline crosswalk

• ISO/IEC 42001:2023 Annex A.6.2.5 — Security of AI systems.
• NIST AI RMF MEASURE 2.7 — Security and resilience evaluated and documented.
• NIST AI RMF MANAGE 4.3 — Incidents and errors communicated to relevant AI Actors.

Penalties

Article 101 fines for GPAI obligations — up to €15 million or 3% of worldwide annual turnover, whichever is higher.

---

Disclaimer. Reference; not legal advice. Verify with counsel. Reg text from Regulation (EU) 2024/1689.