A tiny X post, but a surprisingly interesting one if you build with Claude or Claude Code. It points to a weird, very practical question: when you hit Claude through an API in Chinese and ask “who are you?”, why might it answer as if it were Qwen?
What strikes me is how fast these “identity tells” become folklore in the LLM world. A model saying “I’m Qwen” when probed in Chinese could mean a lot of things: distillation artifacts, prompt contamination, weak identity conditioning, or just a weird failure mode in the specific API setup. I think people are sometimes too eager to jump straight to “it was distilled from X,” because that’s the spicy interpretation, not necessarily the only one.
Still, if this observation is repeatable, it’s genuinely interesting for developers. It hints at how much model personality is really just surface behavior layered on top of training data and instruction tuning. If I were using Claude Code or any Claude API integration seriously, I’d want to know whether multilingual identity prompts are reliable, because these little cracks matter when you’re building evals, routing logic, or safety checks.
What I’d actually do with this is test it instead of meme it. Try the same prompt across languages, across system prompts, and across different API wrappers. I’d be curious whether the “Qwen” answer shows up only in Chinese, whether it depends on the exact phrasing, and whether it vanishes when you make the system message more explicit. That would tell you a lot more than the joke version does.
The takeaway is simple: model identity is fragile, and that fragility can reveal more about training and alignment than a glossy benchmark chart ever will.
Reference: XユーザーのMax For AI（@MaxForAI）さん
