The Legal Boundary of Knowledge Distillation: In 2026, the Clearest Line Is Terms of Service

1. The clearest boundary: ToS breach
This is the line doing the most real work today. OpenAI says users may not use Output to develop models that compete with OpenAI. Anthropic bars access aimed at building competing products or training competing AI models. Large-scale API calls, fake-account creation, and automated extraction therefore become straightforward contract problems before they become anything else. The resulting risk is concrete: account termination, access bans, damages claims, and injunction requests. So when a company distills a closed model at scale without permission, the dominant present-day view is that it has crossed the clearest enforceable line. OpenAI Anthropic

2. Copyright boundary: weak and gray
Copyright protection for model outputs is comparatively weak. US law centers on human authorship, and the U.S. Copyright Office says copyright does not extend to purely AI-generated material or material lacking sufficient human control over expressive elements. Distillation is also closer to transferring behavior or function than to directly copying code or parameters. That makes classic substantial-similarity theories harder to apply. Fair use and reverse-engineering arguments may also enter the discussion, which is why many legal observers think it remains difficult, at least for now, to say simply that distillation equals copyright infringement. U.S. Copyright Office

3. Trade-secret and unfair-competition boundary
Whether outputs themselves qualify as trade secrets remains disputed. But when extraction involves deceptive accounts, circumvention, or clear rule-breaking, the theory becomes stronger: plaintiffs can argue improper acquisition or improper use. The hardest unresolved question is when an ordinary ToS breach becomes an “improper means” case. That is the largest gray zone, and one that likely depends on future case law.

4. Other relevant boundaries
If access is characterized as unauthorized, the Computer Fraud and Abuse Act can become relevant in both civil and criminal settings. And if governments frame industrial-scale extraction as a national-security or industrial-theft problem, sanctions, export controls, and blacklist measures may follow outside ordinary litigation. By contrast, when the source model is open-weight or clearly licensed for such use, distillation is usually treated as ordinary development work. Cornell LII

So the conclusion is not that distillation itself has been branded illegal. The real boundary depends on whose model is extracted, by what means, and at what scale. The most effective line today is terms of service; copyright is secondary, and trade-secret theory is the next major battleground. Until more case law accumulates, this gray zone is likely to remain.