EU AI Act Article 53 (General-Purpose AI Model Obligations)

Article 53 of the EU AI Act (Regulation (EU) 2024/1689) sets out what a provider of a general-purpose AI (GPAI) model must do before and after placing that model on the EU market — baseline obligations including documentation, training-data summaries, and copyright policy. These requirements, set out in Chapter V, became applicable on 2 August 2025.

The four baseline obligations

Every GPAI provider — commercial or academic, whether the model is released openly or kept proprietary — must, at minimum:

• Technical documentation (Annex XI). Keep current documentation of the model’s development and testing process and evaluation results, available on request to the EU AI Office and national competent authorities.
• Downstream-provider documentation (Annex XII). Provide information and documentation to any AI-system provider who integrates the GPAI model into their own system, sufficient for that downstream provider to understand the model’s capabilities and limitations.
• Copyright-compliance policy. Put in place a policy for complying with EU copyright law, in particular Directive (EU) 2019/790, including honoring rightsholders’ opt-outs from text-and-data-mining reserved under Article 4(3) of that Directive.
• Training-data summary. Draw up and publicly release a sufficiently detailed summary of the content used to train the model, using a template the AI Office provides.

The open-source exemption — and its limits

Providers of GPAI models released under a genuinely free and open-source licence — one that allows access, use, modification, and redistribution, with model weights, architecture information, and usage information all made public — are exempt from the technical-documentation and downstream-documentation obligations above. This exemption does not extend to the copyright-policy or training-data-summary duties, both of which apply regardless of licensing model, and it does not apply at all to a GPAI model classified as carrying systemic risk.

Heightened obligations for models with systemic risk

Article 51 creates a rebuttable presumption that a GPAI model carries systemic risk when the cumulative compute used to train it exceeds 10^25 floating-point operations (FLOPs), or when the Commission designates it as such based on high-impact capability criteria in Annex XIII. Providers of a systemic-risk GPAI model face additional duties under Article 55 on top of the Article 53 baseline: state-of-the-art model evaluation including adversarial testing, systemic-risk assessment and mitigation, serious-incident reporting to the AI Office, and adequate cybersecurity protection for the model and its infrastructure. As of this writing, the 10^25 FLOP threshold is a presumption, not an absolute line — the Commission retains authority to designate a model as systemic-risk on capability grounds even below that compute level, and providers can rebut the presumption with evidence the model does not in fact pose systemic risks.

Why this matters for research institutions

Article 53 does not exempt universities, national labs, or nonprofit research consortia as a category. A research group that pretrains its own foundation model, or substantially fine-tunes and redistributes someone else’s, can itself become a GPAI “provider” for AI Act purposes — with the same documentation, training-data-summary, and copyright-policy duties as a commercial AI vendor. This is a materially different question from the research-specific carve-outs elsewhere in the Act: Article 2(6) and 2(8) exempt AI systems developed and used purely for scientific research from most of the Act, but a GPAI model a research institution releases or makes available beyond that narrow purely-research context does not automatically inherit that exemption. Institutions developing or fine-tuning large models should treat model-release decisions — open-weight or not, internal-only or public — as a point where Article 53 obligations may attach, and should not assume nonprofit or academic status is itself a defense.

A fast-moving regulatory picture

The EU’s 2026 “Digital Omnibus on AI” simplification package, politically agreed in May 2026, defers several AI Act compliance deadlines elsewhere in the Act (notably certain high-risk and transparency provisions), but reporting on the Omnibus has not centered Article 53’s GPAI documentation duties, which took effect on schedule in August 2025 and remain in force. Given how actively this area of the Act is still being clarified through Commission guidance and codes of practice under Article 56, institutions with GPAI obligations should verify current status directly against the Commission’s AI Act Service Desk or the Commission’s own GPAI Q&A rather than treat any secondary summary, including this one, as the final word.

Related terms

See also EU AI Act Article 5 (Prohibited AI Practices), EU AI Act Article 6 (High-Risk Classification Rules), EU AI Act Article 10 (Data and Data Governance), foundation model, and frontier model. For the broader compliance picture, see the guide on EU AI Act obligations and exemptions for research organizations.