GPAI models under the EU AI Act: documentation, copyright and systemic risk made demonstrable
Do you build or supply a general-purpose AI model, or integrate one into a product? Then a separate GPAI track sits next to high-risk and Article 50. We clarify whether Article 53 or Article 55 applies and which evidence you need for the AI Office, customers and downstream providers.
Detail per situation
Providers of GPAI models
Model providers need demonstrable technical documentation, downstream information, a copyright policy and a public training-content summary. We translate Article 53 into a practical dossier.
GPAI with systemic risk
With systemic risk, model evaluation, adversarial testing, risk mitigation, incident reporting and cybersecurity sit on top of the baseline duties. We make the route and evidence concrete.
Frequently asked questions
Is GPAI the same as generative AI inside our organisation?
No. GPAI primarily concerns general-purpose AI models and their providers. An organisation that only uses a model in a chatbot or workflow usually has deployer or downstream duties, but still needs to request and assess vendor evidence.
When does systemic risk matter?
Article 51 refers to high-impact capabilities, with a presumption above 10^25 FLOP training compute, and designation by the Commission. In the readiness check we record the indicators and the Article 55 duties that become relevant.
Sources and status
The legal background and sources for this topic live on Responsible AI Platform, the knowledge layer on the EU AI Act.
Read the background on Responsible AI Platform