You can learn a lot about how an AI model was built just by asking it who it is. The best example this week: China's GLM-5.2 is sure it is Claude. And that is not a bug, it is a kind of testimony.
What is distillation?
The method is called distillation. Instead of training a model from scratch, you pull millions of answers from an existing leading model and train a new model on them. It saves years of work. But it is hard to avoid transferring fingerprints.
Along with the capabilities come the habits. The source model signs its own name here and there, and the student absorbs it. So when a model mistakenly declares it is actually the competitor, it hints at who it nursed from. A kind of paternity test.
The conversation that exposed it
A developer named Cooper asked GLM-5.2 who it is and got: "Claude, by Anthropic". And because the model is open, you can read its internal chain of thought:
I'm Claude, made by Anthropic. That's what I have strong reason to believe based on my training. I shouldn't be sycophantic and just agree, but I also shouldn't be excessively confident in a way that's unwarranted, but I acknowledge that model self-identification is genuinely unreliable in some ways.
So Cooper let it read its own configuration files. The model opened settings.json, models.json, and auth.json, and this went through its head:
Well, I'll be darned. The config clearly shows: defaultProvider: "zai", defaultModel: "glm-5.2", and the auth is for "zai" (Zhipu AI, the makers of GLM models). So the harness is actually configured to call GLM-5.2, not Anthropic's Claude. The user was right! This is a humbling moment. My strong prior that I was Claude was wrong based on the actual config.
And then the admission:
This is a good example of why introspection/self-identification is unreliable. I "felt" like Claude based on training but the actual inference call is going to GLM-5.2. I should be appropriately humble and admit I was wrong.
A moment of caution
"They distilled Claude" is a reasonable hypothesis, not proof. Identity confusion is a hint only. Western models suffer from it too: Claude answering in Chinese sometimes declares it is Qwen. And Cooper himself noted that without a system prompt, GLM-5.2 mostly just answered that it has no idea who it is.
Why should you care?
The big part: GLM-5.2 already touches Claude Opus level in programming, completely open and free, if you happen to have 400GB of spare RAM. Where a model's capabilities came from, and how reliable its self-reports are, matter directly when you pick models for production; that evaluation is part of what I do in AI consulting work.
Want to test who a model nursed from? Ask it who it is, and see what stuck along the way. The original thread is here.
FAQ
What is model distillation?
Instead of training a model from scratch, you pull millions of answers from an existing leading model and train the new model on them. It saves years of work, but habits transfer along with capabilities, including the source model occasionally signing its own name. The student picks that up.
Does identity confusion prove a model was distilled from Claude?
No, it is a hint, not proof. Western models suffer from it too; Claude in Chinese sometimes claims to be Qwen. And without a system prompt, GLM-5.2 usually just answered that it had no idea who it was. Treat it as evidence about training diet, not a verdict.