Glossary · EU AI Act: general-purpose AI and transparency
General-purpose AI model (AI Act)
Also known as: GPAI model, foundation model, GPAI model with systemic risk
Under the EU AI Act, a general-purpose AI (GPAI) model is an AI model, including one trained with a large amount of data using self-supervision at scale, that displays significant generality, is capable of competently performing a wide range of distinct tasks and can be integrated into a variety of downstream systems or applications. Models used only for research, development or prototyping before being placed on the market are excluded.
- AI regulation
- EU
- Technical documentation
In one sentence
A general-purpose AI model under the AI Act can perform a wide range of tasks, like a large language model; its providers have documentation duties.
Example
A large language model offered through an API is a GPAI model; a customer-service chatbot built on it is an AI system whose provider relies on the model provider’s documentation.
How it applies
- All GPAI models: Providers must draw up technical documentation of the model (Annex XI) for the AI Office, provide information and documentation to downstream providers who integrate the model (Annex XII), put in place a policy to comply with EU copyright law and publish a sufficiently detailed summary of the training content, using the Commission's template. Open-source models are exempt from the first two duties unless they pose systemic risk.
- Systemic risk: A model is presumed to have systemic risk when its training used more than 10²⁵ floating-point operations. Its provider must additionally evaluate the model, including adversarial testing, assess and mitigate systemic risks, report serious incidents and ensure cybersecurity.
- Dates: The duties apply since August 2, 2025; models placed on the market before that date must comply by August 2, 2027.
- Technical documentation: The documentation for downstream providers is the model's “instructions for use”: capabilities, limitations, acceptable use and integration requirements. Downstream providers need it to document their own systems.
Compared with the USA, Canada and China
California's SB 53 regulates frontier models trained with more than 10²⁶ operations — a threshold ten times higher — and requires published safety frameworks and incident reports. At federal level the USA revoked the reporting duties of Executive Order 14110. Canada's voluntary code covers similar ground without binding force. China regulates the services built on models: a generative AI service must file and pass a security assessment before launch.