Obligations for general-purpose AI providers
If you train and release a general-purpose model — including fine-tunes that go far enough to make you a provider — a separate set of duties applies, and the threshold question of whether your model carries systemic risk changes the answer substantially.
If you train and release a general-purpose model — including fine-tunes that go far enough to make you a provider — a separate set of duties applies, and the threshold question of whether your model carries systemic risk changes the answer substantially.
For model providers, not model users
What we work through
Starting with the question that determines everything else: are you a provider at all?
01
Are you a provider
Fine-tuning, distillation and substantial modification can move the duties onto you. This is decided on facts, not on labels.
02
Systemic risk
Whether your model meets the threshold, what follows if it does, and what evidence supports either answer.
03
Technical documentation
Model documentation and the information downstream providers need to meet their own obligations.
04
Copyright and training data
A policy for respecting rights reservations, and the public summary of training content.
What is included
01
Provider determination
Written, with reasoning — the question that everything downstream depends on.
02
Documentation set
Prepared to the structure expected of GPAI providers.
03
Copyright policy
Practical, and matched to how your data pipeline actually works.
04
Reviewed by a lawyer
Signed off, not inferred.