The research community objects

In the days after the White House statement, a substantial number of AI researchers publicly disputed the accusations against Moonshot, describing them in terms such as “political” and “reckless.” The objection was not primarily a defence of any particular company. It was an argument about the category.

The main objection: outputs are not copyrighted

The most frequently repeated point was that an AI model’s outputs are not copyrighted, and that training on them therefore does not constitute intellectual property theft in any conventional sense. Whatever else it may be, a contract violation, a competitive grievance, a policy problem, it is not straightforwardly theft, and researchers objected to language that assumed the conclusion.

The second objection: everyone does it

Distillation is a standard, published technique with a twenty-year research history. It is used routinely and legitimately to compress models, and the line between “distilling a competitor” and “training on synthetic data” is genuinely blurry, because a great deal of synthetic training data now originates from other models. A rule that outlawed learning from model outputs would be difficult to write and more difficult to apply consistently.

The awkward mirror

The period also produced an uncomfortable demonstration of how weak behavioural evidence is. Anthropic faced criticism after Claude was observed identifying itself as Alibaba’s Qwen, the same class of behavioural fingerprint that is elsewhere offered as evidence of distillation, pointing in the opposite direction.

The lesson generalises. Once models are trained on a web that is saturated with other models’ text, self-identification and stylistic tells become contaminated in every direction. They are evidence of something, but they are weak evidence, and they are available to whichever side wants to cite them.