General-purpose AI, explained for deployers

General-purpose AI, or GPAI, means models with broad capabilities that can be put to many different uses and built into many different tools. Large language models, the engines behind chat assistants, are the obvious example, but image and audio models count too.

They are the foundation a lot of the AI you use sits on top of. The chatbot on a supplier's website, the assistant in your office software, and a bespoke tool a vendor sells you may all be different products built on the same handful of underlying models.

The GPAI rules fall mainly on the model providers, the companies that build and release the models. Their duties include keeping technical documentation, being transparent with the businesses that build on the model, publishing a summary of the training data, and having a policy to respect copyright.

The most capable models, those judged to carry systemic risk, take on extra duties around evaluation and risk mitigation. All of this applies from 2 August 2025, earlier than the high-risk rules, but it is the model maker's burden, not yours.

If you use a product built on a general-purpose model, you are almost always a deployer of that product, not of the model itself. Your duties come from how you use the tool: its risk tier, the transparency duties of Article 50 if it talks to people, and AI literacy under Article 4.

In other words, the GPAI chapter is mostly background for you. You do not need to audit the model maker; you need to classify your own use of the tool that is built on their model.

The picture changes if you go beyond using a tool. Building on a model and offering the result to others, fine-tuning a model into something you provide, or running your own model can move you onto the provider side, with the heavier obligations that follow.

Fine-tuning for your own internal use is generally still deployment; fine-tuning and then selling or distributing the result is where provider duties start to bite. If you are heading that way, get advice early.

For most businesses, GPAI is the engine under the tools you use, and the engine is the maker's responsibility. Your job is the same as with any AI you deploy.

Classify each use, give your people the literacy to use it well, and add transparency where Article 50 applies. The fact that a tool runs on a powerful general-purpose model does not change that checklist.