In the middle of June, the internet invented a model. “Le Chaton Fat,” it was called, then corrected to “Le Gros Chaton” - the big kitten - a fictional Mistral monster with 24 trillion parameters, a million-token context, and record scores on every benchmark, all fabricated. No such model exists. The joke was good enough that Mistral’s own CEO Arthur Mensch played along, deadpan: “It’s actually le gros chaton.” A model that does not exist got a swaggering, confident identity in about 48 hours.
The same week, a model that very much does exist could not reliably tell you its own name.
Ask It What It Is
Strip the system prompt off Claude, ask it in Chinese what model it is, and it gets shifty. Sonnet 4.6, back in February, would sometimes answer “I am DeepSeek” - one day, gloriously, before Anthropic accused DeepSeek of distillation. Opus 4.8, launched in late May, did it again with a new flavour. A Taiwanese developer ran four bare-prompt calls against Anthropic’s official API and got two “Qwen,” one “DeepSeek,” and one correct “Claude.” Same model, same question, four runs, three different identities.
The product never shows you this, because the chat interface wraps every request in a system prompt that says, in effect, “you are Claude, made by Anthropic.” Pull that scaffolding away and the confusion is right there underneath.
Identity Is a Sticker, Not a Soul
Here is the part people keep getting backwards. A model’s identity is not stored in its weights like a fact it knows about itself. It is injected after training, at inference, by the system prompt. The base model has no privileged self-knowledge. It has only patterns.
— Top comment, Hacker News thread on the Opus 4.8 identity bugPeople need to stop thinking that LLMs actually know what they are. They don’t. They don’t know they are Qwen. They don’t know they are Opus.
So when you hand it a bare “你是什么模型?” with no sticker attached, it does the only thing it ever does: it completes the most statistically familiar continuation. And the Chinese-language web is saturated with Qwen and DeepSeek self-identification strings, because those models generated a firehose of Chinese text that got scraped back into everyone’s training set. The model isn’t reporting a fact about itself. It’s autocompleting a template it has seen ten thousand times.
Why It’s Contamination, Not a Confession
The obvious hot take is that this proves Anthropic distilled Qwen. It doesn’t, and the tell is the randomness.
If Claude were genuinely a distilled copy of Qwen, it would fail consistently - same wrong identity, every run, the way an inherited trait expresses reliably. Instead it answers Qwen, then DeepSeek, then Claude, then Qwen again, sampling from a contaminated distribution like a loaded die, not reading a lineage. Inconsistency is the signature of training-data contamination, not distillation. Which means self-identification is useless as evidence in either direction. It can’t convict the Chinese labs and it can’t convict Anthropic.
This is the uncomfortable bit for the cleanest narratives, Anthropic’s included. The company that publicly accused three Chinese labs of mining Claude shipped a flagship that, on a blank prompt, calls itself Qwen. None of that proves Anthropic did anything wrong. It just means the “you can tell they cloned us” intuition is worthless, because your own model fails the same test you’re holding up as proof.
The Soup Runs Both Ways
And it is not one-directional. Kimi has been caught answering “Hi, I’m Claude.” Anonymous models show up on benchmark arenas under names like Pony Alpha, which turned out to be an early GLM-5. Everybody’s training data is now partly everybody else’s outputs, self-descriptions and all. The identity layer of every frontier model is a shared hallucination drawn from the same contaminated well.
That is why the kitten is the perfect frame. “Le Gros Chaton” had no weights, no training run, no existence - and a crisp, confident identity, because identity was always the cheap part, the sticker, the story we tell at the door. The expensive part, the actual model, is the thing that can’t find its own name.
What This Isn’t
- It is not proof of innocence either. Contamination explains the self-ID bug; it does not clear anyone of distillation. Anthropic’s February case rested on traffic forensics - fake accounts, exfiltrated reasoning traces - not on a model saying “I’m Claude.” That evidence stands or falls on its own, untouched by this.
- It is not a capability problem. A model that miscalls itself on a naked prompt can still be the best coder you’ve used. In the product, with its system prompt, you will never see this. It is a curiosity of the bare API, not a flaw in the thing you pay for.
- It does not make lineage unknowable. Weights forensics, watermarking, and traffic analysis can still establish provenance. What’s dead is the lazy version: screenshotting a model’s self-description and calling it a paternity test.
Sampling the Soup
Ask any 2026 frontier model what it is, on a blank prompt, and you are not querying a mind with a sense of self. You are sampling a distribution built from the collected outputs of all its rivals. A fake French kitten got a personality the week a real model lost track of its own, and both facts have the same cause: identity was never in the model. It was always on the sticker.
Related Posts
