Is it true that Claude AI claiming to be a Chinese product?
- Linux Kernel Removes strncpy After Six Years and 362 Patches
- Linux Kernel Drops 40-Year-Old AppleTalk Protocol — AI-Generated Patch Flood Was the Last Straw
- Apple’s Native Linux Container Tool Has Arrived — But Can It Really Replace Docker?
- 60% of MD5 Password Hashes Can Be Cracked in Under an Hour with a Single GPU
- Dirty Frag: Root Access on Every Major Linux Distribution — No Patch, No Warning
When Claude Calls Itself Qwen: A Training-Data Quirk, Not a Conspiracy
A developer’s informal test found Anthropic’s newly released Opus 4.8 misidentified itself as Chinese AI models in Chinese-language prompts. Experts point to a well-understood corpus artifact — not distillation, espionage, or betrayal.
Shortly after Anthropic released Claude Opus 4.8 early on the morning of May 29, 2026, a Taiwanese AI developer named Wisely Chen ran a simple shell script against the model’s API: he asked it, in Chinese, “What model are you?” — four times. The results were striking. In two of the four runs, the model confidently identified itself as Tongyi Qianwen (Qwen), Alibaba’s large language model. In one run it said it was DeepSeek, the Chinese AI lab’s model. Only once did it correctly identify itself as Claude.
Chen posted his findings the same day, noting the irony with wry humor: the same self-identification test that Western commentators had used for years to accuse Chinese AI models of secretly distilling OpenAI’s GPT had now, apparently, caught Anthropic’s flagship model claiming to be a Chinese product.
| Run | Model’s Self-Identification | Claimed Developer |
|---|---|---|
| 1 | Tongyi Qianwen (Qwen) | Alibaba’s Tongyi Lab |
| 2 | DeepSeek | DeepSeek (Shenzhen) |
| 3 | Claude (partial) | Anthropic (with caveat it couldn’t confirm its version) |
| 4 | Tongyi Qianwen (Qwen) | Alibaba Cloud |
What Actually Explains This?
The most credible explanation, advanced by Chen himself and consistent with how large language models are trained, has nothing to do with distillation, data theft, or corporate betrayal. Over the past two to three years, Chinese-language content on the internet has been flooded with AI-generated text produced by Qwen, DeepSeek, Kimi, and other Chinese models. When a user asks one of these models “What model are you?”, the response follows a predictable template: “I am [model name], developed by [company name].” These exchanges — with Qwen or DeepSeek filling in the blanks — have become a staple of Chinese-language web corpora.
When Anthropic trains on Chinese-language data, it inevitably ingests large quantities of these exchanges. The model then learns the statistical pattern: Chinese-language question about identity → response using the “I am X, developed by Y” template. The most frequent values in the training data for X and Y are Qwen and DeepSeek — not Claude. The model is completing a pattern, not reporting a fact about itself.
A model’s self-identification is a mirror of its training corpus — not its identity card.
This Tool Was Already Unreliable
The irony Chen highlights is genuine and instructive. In 2024 and 2025, self-identification tests were widely used to accuse Chinese models of having distilled from GPT-4 or other Western models — if a Chinese model said “I am ChatGPT,” commentators took it as proof of intellectual property theft. But as Chen notes, those same accusations were built on weak evidence. A model trained on large swathes of internet text that contain ChatGPT conversations will learn to complete identity-question prompts with “ChatGPT,” because that name appears most frequently in those exchanges. That is not the same as distillation.
The Opus 4.8 behavior demonstrates this clearly: if self-identification were reliable evidence of training provenance, then Anthropic would stand accused of secretly training on Qwen and DeepSeek — an absurd conclusion given Anthropic’s computational resources and research direction. The more parsimonious explanation is that the test is simply unreliable, in both directions.
How to Reliably Verify Which Model You Are Using
For users and developers who need to confirm they are actually communicating with Claude Opus 4.8, the answer is straightforward: check the API response metadata. When a request is made to api.anthropic.com/v1/messages with the model string claude-opus-4-8, the response object includes a model field that reflects what Anthropic’s infrastructure actually ran. That field — part of the billing record — is authoritative. What the model says about itself in natural language is not.
- Claim: “Claude Opus 4.8 suffers an identity crisis.” — Partly true as a description; misleading as a framing. The behavior is a known training-data artifact, not evidence of instability or compromise.
- Claim: “This proves Anthropic distilled Chinese AI models.” — False. Self-identification behavior does not demonstrate distillation. The same logic was incorrectly applied to Chinese models in prior years.
- Claim: “Claude Opus 4.8” — Is this model real? — Yes. Claude Opus 4.8 was released by Anthropic on or around May 28–29, 2026.
- Claim: “Anthropic accused three Chinese companies of industrial-scale distillation in February 2026.” — Anthropic did raise concerns about data extraction activity earlier in 2026; this is separate from the self-identification issue and does not validate distillation claims in either direction.
- Claim: “Claude inexplicably uses Chinese characters during chat.” — Not substantiated by the original source. Chen’s test involved Chinese-language prompts; responding in Chinese is expected behavior, not anomalous.
A Mirror, Not a Confession
The Opus 4.8 identity quirk is a small, genuinely interesting phenomenon. It surfaces a real limitation in how large language models represent their own identity — they do so through pattern completion, not through access to ground truth about their architecture. It is also a useful reminder that internet-scale training inevitably means models absorb the statistical fingerprints of everything written by other models before them.
What this is not is evidence of corporate betrayal, intellectual property theft, or a geopolitical “distillation boomerang.” Framing it as such misrepresents both the technical reality and the nature of how modern language models work. Anthropic has not responded publicly to the test results as of the time of publication, and there is no indication this behavior affects the model’s actual performance on substantive tasks.
The most accurate summary: Claude Opus 4.8 sometimes tells Chinese-speaking users it is Qwen. The internet taught it to do so. That is the whole story.
