GPAI disclosures: the Art. 53/55 duties
For general-purpose model providers: the Documentation domain holds the downstream model documentation, the copyright policy and the training-data summary — and the extra Art. 55 layer when a model carries systemic risk.
Obligations, filtered to the Documentation domain, collects the duties that attach to general-purpose AI models — Articles 53 and 55 — across your portfolio, beside the Art. 11 technical-documentation and Art. 12 record-keeping duties of ordinary systems. For what counts as GPAI in the first place, and whether these duties are yours at all, start with What counts as GPAI, and what you must do; this article is about doing the work in Veritome.
Which systems appear
Systems where you answered yes — we develop or release a GPAI model on the Model screen of classification. Using someone else's model generates no Chapter V duty; that belongs to the model provider.
The Art. 53 set
- Technical documentation of the model (Art. 53(1)(a), Annex XI) — kept up to date and provided to the AI Office and national authorities on request: training and testing process, evaluation results.
- Information for downstream providers (Art. 53(1)(b), Annex XII) — what an integrator needs to understand the model's capabilities and limitations and to meet their own obligations.
- Copyright policy (Art. 53(1)(c)) — a policy to comply with Union copyright law, including the reservation of rights under the text-and-data-mining exception. It makes a natural evidence record — see The Evidence hub.
- Training-content summary (Art. 53(1)(d)) — the public summary, in the AI Office's template.
Models released under a free and open-source licence get relief from the first two (Art. 53(2)) but not from the copyright policy or the training summary — the checklist reflects that split rather than waving the whole article, and the relief does not extend to models with systemic risk.
The Art. 55 layer — systemic risk
If a model crosses the systemic-risk threshold (the presumption at more than 10²⁵ FLOP of training compute, or Commission designation — Art. 51), a second layer appears automatically: model evaluation including adversarial testing (Art. 55(1)(a)), assessment and mitigation of systemic risk (b), serious-incident tracking and reporting to the AI Office (c), and cybersecurity protection (d). The engine adds these only when the classification says so; the Assessments required strip on Risks shows the Art. 55(1) evaluation as owed until it is done.
Timing
GPAI obligations have applied since 2 August 2025. Models placed on the market before that date have until 2 August 2027 to comply (Art. 111(3)); the GPAI rows carry the date that applies to yours.