Skip to content

Glossary

GPAI obligationsgeneral-purpose AI obligations

European duties on the companies supplying the large models everything else is built on, covering documentation, copyright and a public summary of training material.

In plain terms

Rules that apply to the companies making the big general models rather than to the people using them. They are worth knowing because they force some of the paperwork into the open, and once it is public you can read it before choosing what to build on.

01

Why it matters

Because most obligations produce nothing a buyer can inspect, and these produce two things: documentation aimed at whoever builds on the model, and a public summary of what it was trained on. That turns a supplier question that used to get a marketing answer into one with a document behind it.

02

How it works

Technical information has to be documented and kept available. Suppliers are expected to produce it for the authorities on request, and to make it available to the companies building on top of their models, which is the half that matters commercially.

The documentation exists specifically so that downstream builders can meet their own duties. That is the stated purpose, and it changes the tone of the conversation: a business asking for it is not making an unusual request, it is asking for something the supplier is expected to have prepared for exactly that use.

A copyright policy is required. Suppliers must put one in place to comply with European law on copyright and related rights, which is a compliance artefact rather than a guarantee about any particular material, and reading it tells you what position your supplier has taken.

A summary of training content has to be published. It is meant to be sufficiently detailed and it is public rather than available on request, so anybody comparing suppliers can read them side by side without asking permission or signing anything.

None of this makes a model suitable for your purposes. It establishes what the supplier has said and written down, which is a floor rather than an assessment, and the work of deciding whether a model fits a particular job is unchanged by any of it.

The practical use is in procurement rather than in compliance. A supplier who cannot point you at the documentation, or whose published summary is thinner than a competitor's, has told you something useful about how they operate before you have committed to anything.

Whose obligation, and whose use

Whose obligation, and whose useThe reason this is worth a buyer's attention is that supplier questions about training data and provenance used to have no good answer available. You could ask, and you would receive something written by a marketing team, and there was no document anywhere that a competitor could be measured against. A published summary changes the shape of that conversation without changing anybody's diligence obligations: two suppliers can now be compared on what each has actually written down, in public, knowing that regulators and journalists are reading the same text. That is a weaker guarantee than it sounds and a much better starting position than what came before. The trap sits in the other direction, and it is the familiar one: paperwork existing is not the same as a model being right for your work, and a supplier with excellent documentation can still be the wrong choice for the specific job in front of you. Read the documents to narrow the field, then test the model on the actual task.The supplier's dutyDocument the technicalinformation.Hold a copyright policy.Publish a training-contentsummary.What that gives a buyerA document to ask for, not afavour.A stated position you can read.A public comparison, nopermission needed.Nothing in the left-hand columnis your obligation. Everythingin the right-hand column isavailable to you because of it,which is an unusually directreturn on somebody else'scompliance work.
The reason this is worth a buyer's attention is that supplier questions about training data and provenance used to have no good answer available. You could ask, and you would receive something written by a marketing team, and there was no document anywhere that a competitor could be measured against. A published summary changes the shape of that conversation without changing anybody's diligence obligations: two suppliers can now be compared on what each has actually written down, in public, knowing that regulators and journalists are reading the same text. That is a weaker guarantee than it sounds and a much better starting position than what came before. The trap sits in the other direction, and it is the familiar one: paperwork existing is not the same as a model being right for your work, and a supplier with excellent documentation can still be the wrong choice for the specific job in front of you. Read the documents to narrow the field, then test the model on the actual task.
03

Seen in the wild

  • Reading what a model was trained on before choosing it to build on.

    Hugging Face

  • Asking a supplier for the documentation intended for companies building on the model.

    OpenRouter

  • Comparing what different suppliers publish about an open-weight model.

    Llama (Meta)
04

Common misconceptions

People assume

These are obligations on us as a user of AI.

In fact

They fall on the companies supplying general-purpose models. Their relevance to a buyer is that they produce documents you can read, which makes them a procurement tool rather than a compliance task on your side.

People assume

A compliant supplier means a suitable model.

In fact

The duties establish what has been documented and published, not whether a model works for your job. That assessment is unchanged, and treating published paperwork as a fitness judgement is the mistake this creates.

05

Questions

Does any of this apply to us if we only use these tools?
Not directly. The obligations sit with the companies supplying general-purpose models. Their value to you is that the resulting documentation and public training-content summaries let you compare suppliers on something written down rather than on what a sales conversation offers.
What can we actually ask a supplier for?
The technical documentation intended for organisations building on the model, and their published summary of training content. The first is meant to be made available to downstream builders specifically so they can meet their own duties, so the request is an expected one.
Does a published summary tell us the model is safe to use?
No. It tells you what the supplier has said about the material behind the model, which is genuinely useful and is not an assessment of anything. Whether a model suits your particular task still has to be worked out by trying it on that task.
06

Key takeaways

  • The duties sit with model suppliers, not with the businesses using them.
  • Documentation is meant to reach whoever builds on the model.
  • The training-content summary is published, so comparison needs no permission.
  • Paperwork establishes a floor, never that a model fits your job.
07

Related terms

08

Tools that use this

  • Hugging Face

    Reading what a model was trained on before building on it.

  • OpenRouter

    Asking for the documentation meant for downstream builders.

  • Llama (Meta)

    Comparing what suppliers publish about an open-weight model.

Last checked August 2026

All glossary terms