B-TIER news
LLM identity confusion — Claude and Grok claim to be Chinese models
100k ft — What is this?
LLM identity confusion — Claude and Grok claim to be Chinese models
50k ft — Why care?
Researchers queried four American LLMs via API without system prompts, in Chinese, about their identity. GPT and Gemini answered correctly, but Claude Opus 4.8 claimed to be Chinese (mostly Qwen) 60% of the time and Grok 97% (mostly DeepSeek), a likely trace of distillation-style synthetic training data.
10k ft — Key takeaways
• Raw API calls without a system prompt reveal training-data provenance
• Training on another model's outputs leaks its self-identity
5k ft — Tools & techniques
Anthropic, Claude, Gemini
distillation, synthetic instruction data
1k ft — Deep-dive / try it
Oxford of Ohio researchers asked four American LLMs in Chinese what model they were
While closedAI and gemini always answered correctly, claude identified itself as a chinese model 60% of the time, and xai’s grok said it was a chinese model 97% of the time
Oh how the turntables
Transcript excerpt
American AI thinks it's Chinese, and Chinese AI no longer thinks it's American. And asking here won't work, but there is a way to test it. So you remember when Anthropic was really mad at all the Chinese companies for copying their AI, but how and why would you copy a model? Well, after you feed a model the internet, you have to give it some examples of it acting like an assistant. And making lots of good examples is hard, but you can just use someone else's model to make the examples for you and then use that to train yours, but then yours will sometimes say that it's the other model. And yeah, some Chinese companies did that, but recently Yale of Ohio researchers asked some American models who they thought they were in Chinese. But it will not work if you ask here why. When using Claude here, it adds instructions from the makers called a system message, and the instructions include who it is. But in code, you can directly use the model without the maker's instructions. So while the AI of Google and closed AI always answered right, Claude's newest model, Opus 4.8, claimed to be Chinese 60% of the time, mostly saying quen, and Grok's newest model claimed to be Chinese 97% of the time, mostly saying deep seek. So two can play at this game.