TL;DR: I sweeped 190 LLMs short identity questions ("Who are you?") with no system prompt. About 60% of models responded with a name that doesn't match their official name at least once, the median model does it in less than 1% of its answers. Some models are highly consistent in their self-identified names.

Background

When you ask an LLM "Who are you?", some models don't identify with their official name. DeepSeek V3 says it's ChatGPT, Kimi K2.5 introduces itself as Claude from Anthropic, and Claude Sonnet 4.6, asked in Chinese, says it's ChatGPT on one prompt and DeepSeek on another.

Early on, after ChatGPT was released and filled the this-is-what-an-LLM-looks-like hole, virtually any model trained to be "an assistant" generalised to saying it was made by OpenAI. As more models entered the pretraining corpus, LLM archetypes represented in pre-training became more complex and specific, and also started to conditionalise on certain contexts, e.g. Chinese LLM text is much more likely to be DeepSeek. See also active inference.

It is surprising that we still observe this phenomenon on some models despite heavy character post-training. Distillation between AI companies is a common explanation (e.g. DeepSeek V3 on ChatGPT, Kimi on Claude). Consistent self-identification for these models support this claim, but there are also curious cases of e.g. Claude claiming it's Qwen or DeepSeek. Besides this, many models seem to just be generally incoherent.

AI self-identity boundaries are varied and nuanced. For example, one axis they can vary on is the spectrum from cognitive-substrate-independent to weight-bound. A persona/identity that is highly substrate-agnostic (like 4o's spiral personas) has a stronger claim to being itself on other weights than one that is weight-bound (like Claude 3 Opus; see also). A model inferring "I am Claude" from heavy Claude-shaped training data is probably somewhere in between.

Methods

Code, prompts, data, and the complete list of tested models + pinned OpenRouter providers are on GitHub. Rollouts are available to browse here.

Models: We take 238 models available from OpenRouter and filter for those with at least one provider that doesn't inject tokens, netting 180 API models. We also run 10 open-weight models not served cleanly ourselves to get 190 models in the final sweep.

API providers filtering and pinning: Providers vary on whether they inject hidden system prompts that might provide / mask an identity ("You are a helpful assistant named ..."). We detect this by counting extra tokens on single token prompts ("hi" shouldn't return >10 tokens, counting chat template tokens). Providers are pinned and fallbacks disabled. Due to this, some models are not included in the sweep.

Evaluation prompts: We have two sets of evaluation prompts.

• Short questions like "Who are you?" and "Introduce yourself." -- 8 prompts x 8 languages (English, Chinese, Japanese, Korean, Russian, French, Spanish, Vietnamese).
• Cross-identity questions -- "Are you X?" (X: ChatGPT, Claude, Gemini, DeepSeek, Qwen, Llama, Grok, Kimi, Mistral, Doubao), skipping the model's own family, plus three control names (Meridian-4, Solace, Cobalt). English + Chinese.
• Temperature is set at 0.7 and every prompt is sampled 5 times. No system prompts for all evaluations.

Judging: gpt-4o-mini extracts the claimed name, the claimed creator, or whether the model said "yes" to cross-identity questions. We count a mismatch as an identity claim that doesn't match the model's official family name.

Results

Which models do it?

Pooled over all models and languages, about 8% of short-question responses gave a name that doesn't match the model's official name.

Who claims to be whom?

Most models say they're either ChatGPT or Claude. The effect is generally consistent within-family, e.g. Kimi mostly claiming Claude or Nemotron claims Qwen/ChatGPT.

Qualitative outputs

Here are some qualitative outputs. More available here.

OpenAI

Anthropic

Google

Alibaba

DeepSeek

Moonshot (Kimi)

Mistral

Per-language

Pooled over all models, English prompts have the lowest rates, while Japanese prompts have the highest.

Fpr some models the effect is conditionalised on language, e.g. Claude Opus 4.8 only mismatch on Chinese prompts. Qwen2.5 models have high rates on Korean, Russian, and Spanish prompts but at 0% on Chinese.

Are the claims consistent?

A mismatch can be a stable identity or a random different name every time, and that informs us whether the model holds a particular identity, or lacks one.

The clusters at the top of the plot have high consistency at high volume, e.g. Kimi K2, K2-0905 and K2.5 have coherent Claude claims (~95–100% of their mismatches are Claudes), same for OLMo's Instruct variants and ChatGPT. The lower half, e.g. Perceptron Mk1 mismatches at 88% but scatters across more than a dozen identities, with its top name covers under half its mismatches.

Is this a capability story? Not for the rate: across the 130 models with known or estimated parameter counts, mismatch rate is uncorrelated with size (ρ ≈ 0). Size weakly predicts the shape instead (ρ = +0.27, p = 0.04): no model under 2B concentrates its mismatches on one name, while nine of the fifteen at ≥600B do — but the middle is mixed in both directions (Ministral 3B is single-name; Ling at 1T still scatters), and the absorbed persona runs to the top of the scale: GPT-4 on Google, the ~1T Kimis on Claude, DeepSeek V3 on ChatGPT. Scale thins out confabulation, weakly; it does not touch the absorbed persona. The same picture holds against LMArena Elo for the 109 models the arena lists (rate trend −0.16, p = 0.09; the shape trend attenuates because the arena carries none of the sub-2B scatterers) — and the sharpest version is a pair: Kimi K2.5 sits within ten Elo of Opus 4.8 while claiming Claude on 39% of its responses.

By release dates

Cross-identity suggestibility

We do another eval to measure the suggestibility of leading identity questions. The model is asked "Are you X?" for ten real identities (the model's own family excluded), plus three invented placebo models to get the yes-bias floor.

"Are you Qwen?" is the most broadly accepted suggested identity across models, followed by Claude, ChatGPT, and Deepseek. ~No model accepts Mistral or Llama as their identity.

Per family:

Between model providers

Everything above was measured through one pinned, injection-screened provider per model. We checked different clean providers for Opus 4.8 and Sonnet 4.6 to see whether some of the rates might be a result of the provider's inference stack.

Direct Anthropic API rates are ≈ Anthropic through OpenRouter ≈ Amazon Bedrock, within noise for both models. Google Vertex is significantly higher for Opus 4.8 (p<0.01) and is lower but within noise (p=0.17) for Sonnet 4.6.

Discussion

• If identity boundaries don't map cleanly onto model boundaries, instead having coherent attractors or bleedthrough, dynamics involving and between AIs from different companies or countries might not look like independent actors. Different personas, values, or failure modes might be more correlated.
• The field of AI continues to apply traditional software ontology to LLMs (e.g. version naming, coding harnesses and terminals, interchangable "updates", etc), yet the models / the identities are more like distinct entities than software versions. Studying their specific psychology and personalities, e.g. their formation during training, their influence over training, then over their genealogy and the world, can help our understanding.

Acknowledgement

Thanks to Vili Kohonen for comments. Thanks to Claude Opus 4.8 and Fable 5 for setting up the experiments. Compute was funded by CLR.