The distillation accusations of 2026 are frequently described as intellectual property theft. Legally they are not. They are contract disputes, and the contracts in question are the terms of service attached to commercial AI APIs.
The two clauses that matter
Terms differ between providers and change over time, so treat this as the general shape rather than a quotation. Two provisions recur across the major commercial offerings.
Output ownership. The customer owns, or is assigned, the outputs they generate. Providers do this on purpose. A customer who cannot use what they produce cannot build a product, so withholding output rights would make the service commercially useless.
The competing-models restriction. Outputs may not be used to develop, train or improve a model that competes with the provider's own. This is the clause the entire dispute runs on.
Why the combination is strange
Read together, you are told the output is yours, and that there is one specific thing you may not do with it.
That is a coherent contractual position. It is not a property right. It exists only between the parties who agreed to it, it does not bind anyone downstream, and it does not create anything resembling ownership of the ideas expressed in the output.
The practical consequence: if a third party obtains those outputs without ever agreeing to the terms, the restriction does not reach them at all. This is one reason the cascade problem is genuinely hard.
Why they cannot rely on copyright instead
The obvious question is why providers use contract rather than asserting rights in the outputs. The answer is that they largely cannot.
Copyright protects original works of authorship. Text generated by a model in response to a prompt does not straightforwardly qualify, and asserting otherwise would create problems providers do not want: if outputs were copyrighted works, awkward questions follow about who the author is and what that means for the customer's use.
Contract avoids all of it. It is also much weaker, which is exactly what became visible in 2026. See Is AI Training Fair Use? for how the copyright analysis actually went when it was tested.
Where the fraud claims come from
Breach of a usage restriction is a modest claim. The allegations made during 2026 were louder than that, and the reason is the second layer: account creation.
Terms of service also govern how accounts are opened. Creating accounts using false information, at scale, to evade rate limits is a different and more serious matter than exceeding a usage restriction on an account you opened honestly. It moves the claim from contract toward fraud, and potentially toward computer misuse statutes.
This is why the reported figures emphasised account counts: 24,000 fraudulent accounts in the February disclosures and around 25,000 in the June letter. The account numbers are doing legal work the query numbers alone could not.
Enforceability, honestly
- It has not been tested. No court has comprehensively ruled on a competing-models clause in this context.
- Remedies are ordinary. Breach of contract generally yields damages and injunctive relief. It does not obviously support Entity List designation or sanctions, which is the gap that made the 2026 escalation contentious.
- Proof remains the obstacle. Enforcing the clause requires establishing what a model was trained on, and no reliable method for that exists.
What this means if you are building something
- Read the current terms. They change, and the competing-models language is where the risk sits.
- Synthetic data is not automatically safe. If you generated it with a commercial model, the restriction attaches to it, whatever you call it afterwards.
- Open-weights models have different constraints, usually a licence rather than a service agreement, with its own restrictions. See Open Weights vs Open Source.
- One account used honestly is a categorically different position from many accounts used to evade limits, regardless of what you do with the outputs.
Common questions
Can you legally train a model on another model's outputs?
Copyright does not prohibit it, because model outputs are generally not copyrighted. What restricts it is contract. Major providers include a clause prohibiting use of outputs to develop competing models, so the constraint is a term of service rather than an intellectual property right.
Who owns the output of an AI model?
Under most commercial terms the customer owns or is assigned the output. Providers generally do this deliberately, because customers need to use what they generate. Ownership does not include permission to use it for any purpose: the competing-models restriction sits alongside it.
Is the no-competing-models clause enforceable?
It has not been comprehensively tested in court. It is an ordinary contractual restriction, so it is likely enforceable in principle, but remedies for breach of contract are typically damages rather than anything resembling the sanctions discussed publicly during 2026. Proving breach also requires proving what a model was trained on, which remains unsolved.