Does a Model Calling Itself Claude Prove It Was Copied?

Moonshot's Kimi K3 shipped this week and went straight to the top of Arena.ai's frontend coding leaderboard, beating Anthropic's Fable 5 on pairwise win rate. Within a day, a screenshot was circulating in which the model, asked what it was, answered that it was Claude, made by Anthropic. Wccftech ran it under a headline about the model betraying its distilled origins, and the argument wrote itself: the Chinese model says it is Claude, therefore it was trained on Claude.

It is a satisfying piece of evidence. It is also close to worthless, and it is worth being precise about why, because there is a real version of this argument and this is not it.

Why models get their own names wrong

A model has no privileged access to its own training history. It has no memory of being trained and no internal record it can consult. When you ask what it is, it does what it does with every other prompt: it predicts a plausible continuation based on patterns in its training data.

And its training data is the internet, which by now contains an enormous volume of transcribed chatbot conversation. Overwhelmingly those transcripts feature ChatGPT and Claude, because those are the assistants people have been pasting into blog posts, forums, Stack Overflow answers, and GitHub issues for years. So the text pattern "user asks what are you / assistant answers I am Claude, made by Anthropic" appears in the corpus many thousands of times. A model that has read all of it will complete that prompt the way the corpus completes it, unless someone has specifically trained it not to.

This is not a Chinese-model phenomenon. Gemini has claimed to be ChatGPT. Early Llama fine-tunes did it constantly. Even models from labs with no plausible access to a competitor's outputs have done it, because the artifact comes from public web text, not from any private pipeline. Identity is one of the few questions where a language model is a genuinely unreliable narrator about itself.

The real argument, which is separate

None of which means the distillation concern is invented. In February, Anthropic published a detailed account alleging that DeepSeek, Moonshot, and MiniMax had systematically extracted capabilities from Claude using roughly 24,000 fraudulent accounts across more than 16 million exchanges. The company said DeepSeek's effort targeted reasoning traces, and that Moonshot's focused on agentic reasoning, tool use, and coding.

That is a substantively different kind of claim. It rests on account records and usage logs, evidence Anthropic actually holds, rather than on what a model says about itself. If it is accurate, it describes deliberate, industrial-scale terms-of-service violation.

It is also contested. Beijing has called the allegations groundless. Forbes noted that Anthropic has published blog posts and letters to legislators while releasing relatively little of the underlying technical evidence, and some researchers read the campaign as at least partly a bid for regulatory protection against cheaper competitors. Those two readings are not mutually exclusive. A company can be telling the truth about being copied and also finding that truth commercially convenient.

What this should change

Treat the screenshot and the account logs as separate things, because they are. One is a well-understood corpus artifact that proves nothing about any particular model's provenance. The other is a specific, checkable allegation that deserves to be resolved with evidence rather than by press release.

The reason to keep them apart is practical. If distillation from frontier APIs really is happening at the scale alleged, it is a serious question for how these companies protect their work and how anyone verifies where a model's capabilities came from. Arguing it on the strength of a model that misremembers its own name makes the case easier to dismiss than it should be.

We have looked before at how open-weight models get judged against different standards from closed ones. This is the same pattern in a new form. The strength of an argument should not depend on which flag the lab operates under.