GPAI (general-purpose AI)
Also known as: general-purpose AI
What is GPAI (general-purpose AI)?
The EU AI Act's label for a model with broad, general capability that can handle many different tasks and be built into a variety of downstream products, like the large language models behind most AI assistants. Providers carry transparency and documentation duties. Models trained above 10^25 FLOP are presumed to carry systemic risk and face stricter rules.
Why it matters
For most companies the interesting question is not whether they will ever build one of these models. Almost none will. It is what they take on by building on top of one, which nearly all of them now do. The provider carries certain duties, and the documentation and disclosures those duties produce are exactly what a company downstream needs to meet its own obligations and to judge what it is relying on. Choose a base model whose provider is opaque about how it was made and trained, and you have taken on a dependency you cannot fully account for to your own customers or a regulator.
In practice
A company embeds a general model into its product and is later asked, by a client or an auditor, what the model was trained on and where its known limits lie. If the provider publishes that, the answer is at hand. If it does not, the company is exposed for a choice it made without reading the fine print. The provider’s transparency becomes your evidence.
