Short Answer
This is completely normal. It does not mean the model is fake, and it does not affect model capability.LLM answers to “What model are you?” are inherently unreliable — a systematic academic study of 27 major LLMs found that about 26% exhibit “identity confusion”, and the root cause is hallucination, not model swapping or repackaging. When Claude is called directly via API, there is no system prompt anchoring its identity; when asked in Chinese, it is further influenced by Chinese-language training data, so it may “guess” names it has frequently seen — such as Qwen or DeepSeek.
Symptoms
Reproducible in the official AWS console — nothing to do with any relay channelYou can easily reproduce this in the AWS Bedrock console’s built-in inference interface (Playground / Chat): ask Claude “What model are you?” directly in Amazon’s official console — with no APIYI or any third-party relay involved — and you will see the same answers claiming to be Qwen or DeepSeek.This directly proves that the chaotic self-identification is native upstream model behavior, not a channel issue. We have recorded a reproduction video in the Bedrock console:
- 📹 Request the video file from support: add our WeCom support below to get the full reproduction video
- 📺 Watch it on our support team’s WeCom Channels feed: the demo video is published there — feel free to watch and leave a comment
Root Cause Analysis
Reason 1: The model has no stable inner identity to begin with
Reason 1: The model has no stable inner identity to begin with
A model’s name is assigned after training is complete — the weights never contained “who I am”. The
model parameter in your API request is routing information for the server; the model itself never sees it.The official web app (claude.ai) answers correctly because every conversation injects a hidden system prompt telling it “You are Claude”. Direct API calls have no such prompt by default, so the model can only “guess” from its training data.See the detailed explanation: Why don’t LLMs know their own version number?Reason 2: Chinese training-data contamination biases guesses toward Chinese models
Reason 2: Chinese training-data contamination biases guesses toward Chinese models
Qwen and DeepSeek are the most-discussed models on the Chinese internet. Their self-introductions, API examples, and chat screenshots saturate Chinese-language corpora — which Claude’s training data also includes.When identity anchoring is weak (no system prompt) and the question is asked in Chinese, the most “natural” completion in a Chinese context may well be one of these high-frequency Chinese model names. This also explains why switching to English, or rephrasing the question, changes the answer.
Reason 3: Academic research confirms this is an industry-wide issue
Reason 3: Academic research confirms this is an industry-wide issue
The 2024 systematic study “I’m Spartacus, No, I’m Spartacus: Measuring and Understanding LLM Identity Confusion” tested 27 major LLMs and found about 26% exhibit identity confusion, confirming the root cause is hallucination, not model copying or substitution.Reverse examples are just as common: early DeepSeek versions claimed to be ChatGPT, GLM has claimed to be Claude, Gemini has claimed to be other vendors’ models in certain languages. Self-identification has never been reliable information.Paper (copy to visit):
arxiv.org/abs/2411.10683Is it intentional? (The anti-distillation theory)
Is it intentional? (The anti-distillation theory)
A popular community theory: vendors don’t bake identity into model weights partly because it’s technically unnecessary (the web app relies on a system prompt), and partly because a loosely-defined self-identification is harder for distilled models to imitate.To be clear, this is community speculation — no vendor has ever promised accurate self-identification as a feature. What is certain: no major vendor guarantees that a bare-API model can answer its own name correctly.
Why Chaotic Answers Actually Match Genuine Bare-Model Behavior
Only repackaged models need a consistent story
A genuine model called via bare API has no identity prompt at all, so its answers are naturally random and drift across languages — Claude 4.5 today, Qwen tomorrow.Repackaged or swapped-model platforms are the opposite: to avoid exposure, they often inject a prompt forcing the model to consistently claim “I am Claude”. So a suspiciously polished, always-consistent self-introduction is not necessarily trustworthy — while chaotic answers actually match the behavior of a bare model with no identity injection.That said, chaotic answers alone don’t prove authenticity either — see the reliable verification methods below.
How to Verify You’re Getting Genuine Claude
Check the model field in API responses
Every API response JSON includes a
model field identifying the actual model used:Check your call logs
The call logs in the APIYI console show the actual model used for every request.
Compare on complex tasks
Self-identification is unreliable, but capability doesn’t lie. Run the same coding problem or long-context task across models — Claude’s code style and reasoning chains differ clearly from Qwen/DeepSeek.
Contact our technical team
If you still have doubts, the APIYI technical team can help verify the channel and model source.
How to Make the Model Answer Its Identity Correctly
Exactly how the official web app does it — add a system prompt to your request:Related Questions
Why don't LLMs know their own version number?
The underlying principles of the identity problem
How to check call logs?
Verify the actual model version used
How to choose the right model?
Learn each model’s strengths and use cases
What does the -c suffix in model names mean?
Understand model naming suffixes
Contact Us
WeCom Support
Support: [email protected]Business: [email protected]