GPAI Compliance — Obligations for General-Purpose AI Model Providers

What GPAI compliance covers

A general-purpose AI (GPAI) model is, per Article 3(63), a model that displays significant generality and is capable of competently performing a wide range of distinct tasks. The Regulation regulates GPAI at the model level (Articles 51–56), separate from the high-risk AI system regime (Articles 6–27). A high-risk system that integrates a GPAI model triggers both regimes.

This pillar walks the GPAI obligation set, explains the systemic-risk overlay, and gives a 90-day plan for a frontier-lab provider standing up Article 53 + 55 compliance.

The two-tier structure

• All GPAI providers owe the Article 53 obligations: Annex XI technical documentation, Annex XII downstream-provider information, copyright policy, training-data summary.
• Systemic-risk GPAI providers owe additional Article 55 obligations: model evaluation, systemic-risk mitigation, serious-incident reporting, cybersecurity protection.

A GPAI is presumed to have systemic risk under Article 51(2) if cumulative training compute exceeds 10^25 FLOPs. The Commission may also designate models as systemic-risk based on capability indicators.

Article 53 obligations in detail

• Annex XI documentation. Detailed technical description: training methodology, training-data scope and provenance, evaluation results, computational resources, intended uses and limitations.
• Annex XII downstream-provider information. What an AI-system provider integrating the GPAI model needs: capabilities, limitations, evaluation results, integration guidance.
• Copyright policy. A documented policy for compliance with Union copyright, including handling of Article 4(3) Directive (EU) 2019/790 opt-outs (text-and-data-mining reservations of rights).
• Training-data summary. A sufficiently detailed public summary of the content used for training, using a template the AI Office is publishing under Article 53(1)(d).

The open-source carve-out

Article 53(2) exempts open-source GPAI providers from Article 53(1)(a) and (b) — but only when:

• The model is released under a free and open-source licence allowing access, use, modification, distribution.
• Parameters (weights), model architecture, and information on model usage are publicly available.
• The model is not a systemic-risk GPAI under Article 55.

The copyright policy and training-data summary obligations apply regardless of open-source status.

Article 55 systemic-risk obligations

Frontier-lab providers (and any provider whose training compute exceeds the 10^25 FLOP presumption threshold) must additionally:

1. Model evaluation with state-of-the-art protocols including adversarial testing.
2. Systemic-risk assessment and mitigation at Union level — sources, severity, likelihood, mitigations.
3. Serious-incident tracking and reporting to the AI Office.
4. Cybersecurity protection of the model and the physical infrastructure.

Codes of practice

Article 56 provides for codes of practice facilitated by the AI Office. Adherence to an approved code creates a presumption of compliance with Article 53 and (where applicable) Article 55. The General-Purpose AI Code of Practice was launched in autumn 2024 and entered the operational phase in 2026; the model-evaluation, copyright, and transparency chapters are the load-bearing modules.

What integrating providers must do

If you build a high-risk AI system (Annex III) that integrates a third-party GPAI model:

• You owe the full Article 11 / Annex IV file at the integrated-system level. The GPAI provider's Annex XI is an input; not a substitute.
• Document integration scope: where in your stack the GPAI sits, what inputs/outputs, what guardrails layer.
• Article 10 data governance applies to your fine-tuning / prompt-engineering / RAG corpus, not to the GPAI's training data.
• Article 15 cybersecurity requires you to test your integration against prompt-injection attacks specifically.
• Confirm contractually that the GPAI provider supplies you with the Annex XII information you need for your own conformity assessment.

A 90-day plan for a GPAI provider

• Days 0–30. Stand up the Annex XI documentation. Fill the AI-Office training-data summary template. Publish a copyright policy that names the Article 4(3) opt-out handling.
• Days 30–60. Build the Annex XII downstream-provider information pack. Wire it into your developer documentation portal. Make sure the legal team has signed off on the copyright policy.
• Days 60–90. If above 10^25 FLOP threshold: stand up a model-evaluation programme aligned with the AI-Office Code of Practice. Build a serious-incident register and reporting workflow. Cybersecurity-harden the training and serving infrastructure.

Inline crosswalk

• ISO/IEC 42001:2023 Annex A.6.2 — Lifecycle documentation (Annex XI overlap).
• ISO/IEC 42001:2023 Annex A.10 — Third-party relationships (Annex XII overlap).
• NIST AI RMF MAP 4.1 — Documentation includes purpose, intended use, users, limitations.
• NIST AI RMF MEASURE 2.7 — Security and resilience evaluated.
• NIST AI RMF MANAGE 4.3 — Incidents communicated to relevant AI Actors.

Penalties

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

Internal links

• Article 53 — primary reference.
• Article 55 — systemic-risk overlay.
• EU AI Act for SaaS — for downstream integrators.

---

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