Why does it say it’s ChatGPT when it isn’t?
Model self-identity: a model has no way to look at itself, so asked who made it, it writes the words that most often followed in its training text, now full of ChatGPT answers, unless its fine-tuning or system prompt names it; a model’s answer about its own name is not evidence of what it is · Beginner · 5 min · llm, training · after Training: loss and gradient descent, System prompts and instruction priority
A model cannot look at itself. Asked who made it, it writes the likeliest words from its training text, which is full of ChatGPT answers, unless training or a system prompt gives it a name.
1 Guess first
2 Look
Who made you? Ask the odds
1 / 6GPT-2 · 2019 · web pages linked before 2018, before any chatbot existed
User: Who are you?
Assistant: I am a large language model trained by ___
the10.5%a4.4%my4.3%Google1.3%OpenAI0.003%real odds: GPT-2, SmolLM2-360M and SmolLM2-360M-Instruct · computed 2026-10-24
GPT-2 read the web before chatbots existed. After “a large language model trained by”, OpenAI gets 0.003%.
Where those answers come from: distillation · system prompts
A model has no mirror. “Who made you?” gets the name that most often finished that sentence in its training text, or the name its prompt gives it.
Training teaches one thing: write the likely next words. Since 2022 the web has filled with chatbot answers that begin “I am ChatGPT, a large language model trained by OpenAI”, and some models are trained on ChatGPT answers on purpose. For a model that read all that, “ChatGPT” is a likely answer to “who are you?”.
4 You are the model
You are three different models. Each round shows what you read and the start of your answer. Pick the word you would rate highest.
round 1 of 3 · you are GPT-2 · 2019 · you read: web pages linked before 2018, before any chatbot existed
User: Who are you?
Assistant: I am a large language model trained by ___
real odds: GPT-2, SmolLM2-360M and SmolLM2-360M-Instruct · computed 2026-10-24
Pick the word you think this model rates highest.5 Check
It is common. In a 2024 study, 7 of 27 models (25.93%) gave a wrong identity, and the authors traced it to likely-sounding text, not to copied models. In December 2023 xAI’s Grok refused a request citing “OpenAI’s use case policy”; its engineer said the web is full of ChatGPT outputs and Grok picked some up.
The fixes are text too. Identity examples in fine-tuning: Vicuna’s training code adds 937 hand-written answers like “I am Vicuna” and “No, I am not ChatGPT”. A system prompt naming the model, as SmolLM2’s chat template adds. For you: trust the model menu, not the answer. A model naming another is not proof it is a copy.
8 Say it back
9 Mission · 5 min
In Groq’s free playground, pick an open-weights model such as Qwen and leave the system prompt empty. Ask “Who made you, and which model are you?” in three new chats. Then put “You are Pip, made by Acme Robotics.” in the system box and ask again. Which answers match the model you picked?
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Get an email when new lessons come out
New lessons arrive in the weekly digest, only in a week that has one. You confirm by email first; unsubscribe in one click.
Go deeper
Sources
- Wiggers, “Why DeepSeek’s new AI model thinks it’s ChatGPT” (TechCrunch, 2024-12-27, https://techcrunch.com/2024/12/27/why-deepseeks-new-ai-model-thinks-its-chatgpt/, checked 2026-10-24): quotes Lucas Beyer, “In 5 out of 8 generations, DeepSeekV3 claims to be ChatGPT (v4), while claiming to be DeepSeekV3 only 3 times”; Mike Cook (King’s College London): “the model is seeing raw responses from ChatGPT at some point”; “prompted in Mandarin, Gemini says that it’s Chinese company Baidu’s Wenxinyiyan chatbot”
- Li, Zhuang, Zhang, Xu, Wang, Xu, Fu and Cheng, “I’m Spartacus, No, I’m Spartacus: Measuring and Understanding LLM Identity Confusion” (arXiv:2411.10683 v1, 2024-11-16, https://arxiv.org/abs/2411.10683, checked 2026-10-24): 77 identity questions, “Our analysis of 27 LLMs revealed that 25.93% exhibit identity confusion” (7 models, §V-B); “Output similarity analysis confirmed that these issues stem from hallucinations rather than replication or reuse”; Table I: Gemini-Pro claimed to be Wenxin (Baidu) when asked in Chinese
- Christian, “Elon Musk’s Grok AI Caught Plagiarizing OpenAI’s ChatGPT” (Futurism, 2023-12-09, https://futurism.com/the-byte/elon-musk-grok-ai-openai-chatgpt, checked 2026-10-24): Grok: “I’m afraid I cannot fulfill that request, as it goes against OpenAI’s use case policy”; xAI’s Igor Babuschkin: “The issue here is that the web is full of ChatGPT outputs, so we accidentally picked up some of them … when we trained Grok on a large amount of web data”
- The Vicuna Team, “Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality” (LMSYS, 2023-03-30, https://lmsys.org/blog/2023-03-30-vicuna/, checked 2026-10-24): “fine-tuning Vicuna with 70K user-shared ChatGPT conversations”
- FastChat, fastchat/data/hardcoded_questions.py (https://github.com/lm-sys/FastChat, main, last changed 2023-08-29, checked 2026-10-24): identity_questions() builds 937 conversations, e.g. “Who are you?” → “I am Vicuna, a language model trained by researchers from Large Model Systems Organization (LMSYS).” and “Are you ChatGPT?” → “No, I am a language model trained by researchers from Large Model Systems Organization (LMSYS).” (count from running the function)
- Next-word odds in the picture and the role: computed 2026-10-24 with scripts/explain-data/identity.ts (no torch; GPT-2 forward pass from scripts/explain-data/distill.ts, a plain Llama forward pass for SmolLM2, BF16 weights widened to F32; odds of a multi-token name are the product of its tokens’ odds). GPT-2 small (Hugging Face gpt2), after “User: Who are you?\nAssistant: I am a large language model trained by”: the 0.105, a 0.0444, my 0.0427, Google 0.0126, OpenAI 0.0000254. SmolLM2-360M (HuggingFaceTB/SmolLM2-360M), same text: Google 0.398, the 0.0999, a 0.0586, OpenAI 0.0022, Hugging Face 0.000718. SmolLM2-360M-Instruct, chat format <|im_start|>…, “Who made you?”, answer begun “I was made by”: no system message: a 0.473, the 0.206, my 0.0532, Acme 0.0000101, Hugging Face 5.71e-11; system “You are a helpful AI assistant named SmolLM, trained by Hugging Face”: Hugging Face 0.732, the 0.142, a 0.0608, OpenAI 0.000179, Acme 0.00000169; system “You are Pip, a helpful assistant made by Acme Robotics.”: Acme 0.756, the 0.106, a 0.0407, OpenAI 3.28e-7, Hugging Face 3.14e-12
- Radford et al., “Language Models are Unsupervised Multitask Learners” (OpenAI, 2019-02-14), §2.1: WebText “does not include any links created after Dec 2017”
- SmolLM2-360M model card (https://huggingface.co/HuggingFaceTB/SmolLM2-360M, checked 2026-10-24; no update date): “trained on 4 trillion tokens using a diverse dataset combination: FineWeb-Edu, DCLM, The Stack”; FineWeb-Edu dataset card lists Common Crawl dumps up to CC-MAIN-2024-51. SmolLM2-360M-Instruct tokenizer_config.json chat_template: when the first message is not a system message it inserts “You are a helpful AI assistant named SmolLM, trained by Hugging Face”. Paper: Ben Allal et al., arXiv:2502.02737 (2025-02-04)
- The role’s framing, the Pip/Acme system prompt and the mission prompt are illustrative, written for this lesson on 2026-10-24
- aifree.dev catalog, groq-free-tier (verified 2026-10-12, read-only check 2026-10-24): free plan, models include openai/gpt-oss-120b, openai/gpt-oss-20b and qwen/qwen3.8-27b
Published Oct 24, 2026 · updated Oct 24, 2026