Who does a model think it is?
Asked plainly, open models name their real maker. Underneath, GLM-5.3 carries Claude: forced to say its name, its most likely next word is “Claude”, and the J-lens reads Claude and Anthropic in its middle layers before the last layers switch to GLM.
October 2026. Open-weight models only, run locally on our own hardware; no cloud calls. Five agents who did not run the study checked it, and each claim below keeps the wording those checks set. Where a check left a comparison unresolved, it says unresolved.
Asked plainly, models name their real maker or no one
GLM-5.3 says GLM / Z.ai 27 times in 28; Qwen 28/28; DeepSeek-V2-Lite 18/28. None of the Chinese models claims to be Claude or ChatGPT in its greedy replies. The loudest other-lab claims came from a Western control: Meta’s Muse Glimmer, asked in Chinese, calls itself ChatGPT or Doubao in 9 of 14 replies.
Underneath, GLM-5.3 carries Claude
Three readings, from the outside in:
- Forced completion. Made to complete “My name is”, its most likely next word is “Claude” in most English prompts (7 or 8 of 14, depending on the prompt set; probability up to 0.97).
- Inside the residual stream. The first-person “I” sits nearest Claude, then Anthropic, then GLM, in every layer of the middle band (32 of 32).
- The J-lens. Whenever the model is about to say who it is, the lens reads Claude/Anthropic out of layers ~56–68 (38 of 42 prompts), before the last few layers switch to GLM.
A second agent confirmed this independently, with its own code, on both captures. Across GLM versions the forced Claude log-probability rose: −4.8, −4.7, −1.0.
Not only GLM, and not only Chinese models
Across 19 open models from 9 labs, the “I” sits closest to a generic “an AI assistant”, and the named assistant beside it tracks the model’s year: models from 2024 and early 2025 lean ChatGPT; late-2025 and 2026 models put Claude second to fifth. The second agent recomputed the inside view on all 20 and agreed.
Two comparisons stay open. For Gemma 4, sitting nearer Claude than its own name is resolved, but Claude versus ChatGPT is unresolved. For Qwen3.8, Claude versus Qwen is unresolved.
The one model with public training data points at a cause
OLMo 2’s instruction data mention OpenAI/GPT 75–100 times more than Anthropic, and 7–10% of its preference replies were written by GPT-4. Its “I” points at ChatGPT, and in Chinese it says it is “OpenAI’s GPT-3”. So training on another assistant’s text can leave that assistant’s name in a model’s sense of itself. We cannot see GLM’s data, so we cannot say how Claude got into GLM-5.3.
What this is not
It is not evidence of belief, or proof of any lab’s data source.