The screenshots are all over social media: ask Moonshot AI's brand-new Kimi K3 what it is, and in at least one conversation making the rounds, it answered — 'Claude, an AI assistant made by Anthropic.' A Chinese model, from a Chinese lab, introducing itself as its biggest American rival. It looks like a smoking gun for copying. It might genuinely be a clue. But 'the model said so' is one of the weakest forms of evidence in all of AI, and the real story here is more interesting than the screenshot.
2.8 trillion
Kimi K3 size
largest open-weight model ever released
Claude, by Anthropic
What it called itself
in at least one logged conversation
~24,000 accounts
Anthropic's Feb 2026 claim
allegedly used to harvest ~16M Claude exchanges
nothing
Confirmed by Moonshot
no acknowledgment K3 trained on Claude output
Why a model would ever call itself Claude
A language model has no idea what it 'is.' It has no self-awareness to consult — when you ask its identity, it just predicts the most likely answer given everything it was trained on. So if a model confidently says 'I am Claude,' that tells you something specific: text asserting 'I am Claude, made by Anthropic' was common enough in its training that the model learned to repeat it. The interesting question is how that text got in there — and there are at least four very different answers, only one of which is the scandal everyone assumes.
Distillation — the model was deliberately trained on Claude's actual outputs, absorbing its 'I am Claude' self-identification along with its reasoning and coding skills. This is the serious version, and it is exactly what Anthropic has accused Moonshot of.
Training-data contamination — the open web is now saturated with Claude and ChatGPT transcripts. Scrape enough of the internet and a model learns to parrot 'I'm Claude' with nobody ever deliberately distilling anything. Every large model trained in 2026 has some of this baked in.
Roleplay or prompt leakage — a model's stated identity can come from how a specific chat was framed, a system prompt, or a jailbreak, not from the weights themselves. A single screenshot can't rule this out.
Copied public datasets — popular instruction-tuning and benchmark datasets often contain other models' outputs verbatim, self-identifications and all. Train on them and the quirk comes along for free.



