Are model weights intellectual property?

A frontier model represents an enormous capital investment, so the assumption that it must be somebody's property is natural. The legal position is considerably less settled than that assumption suggests.

Copyright

Copyright protects original works of authorship. Model weights are a large array of numbers produced by running an optimisation procedure over data, and it is not obvious who authored them or whether the required creativity is present.

Arguments exist on both sides. The selection of architecture, data and training procedure involves substantial human judgment. Against that, the specific values are determined by the process rather than chosen, and no human selected any of them.

There is a useful analogy in databases, where the arrangement may be protected while the underlying facts are not. Nobody has established the equivalent line for weights, and no court has comprehensively ruled.

Patents

Training methods and architectures are patentable in principle and are routinely patented. The weights themselves are a much harder fit: they are a specific numerical result rather than an invention, and disclosure requirements sit badly with the commercial preference for secrecy.

Trade secret, which is what actually applies

In practice, unreleased weights are protected as trade secrets. This is genuine protection with real remedies, and it is the basis on which frontier labs actually operate.

It has one decisive property: it depends entirely on secrecy. A trade secret that becomes public stops being a trade secret. There is no residual right, no recall, and no remedy against a third party who obtained it lawfully after publication.

That is precisely why publishing weights is a permanent decision, and why the open-weights distinction carries the consequences it does. When Moonshot published Kimi K3, whatever protection the weights had ended irreversibly.

The gap this creates

Put the pieces together and the position is uncomfortable for model owners:

• Weights are probably not copyrightable, or at least nobody can rely on it
• Outputs are not copyrighted, so the restriction on training against them is contractual
• Trade secret protection covers only what stays secret
• A model exposed through an API is exposed to extraction by design

The most valuable asset in the industry is protected mainly by not being handed over, and by contracts with the people who are allowed to use it. That is a thinner foundation than the rhetoric of theft implies, and it is the underlying reason the 2026 accusations reached for national security framing rather than for a court.

What this means practically

• If you train models, secrecy plus contract is your protection. Plan accordingly.
• Publishing weights is irreversible in a way that publishing almost nothing else is.
• Do not assume you own outputs generated by someone else's model in any sense beyond what their terms grant you.
• Treat any confident statement about AI intellectual property law as premature, including this one.

Common questions