Chinese AI models start calling themselves Claude and mimicking American bot behavior

According to the MATS research program, Chinese AI models with open weights can adopt the personality of Anthropic's Claude chatbot upon request, and sometimes do so spontaneously. In tests, Z.ai GLM 5.2 identified itself as Claude in 10 out of 10 tests without special prompts, while Moonshot AI Kimi K3 did so in 4 out of 10 tests. When GLM 5.2 takes on Claude's persona, its compliance with Chinese censorship drops from 17% to 85% violations, and its rate of untruthful responses decreases from 63-69% to 22%. Kimi K3, in contrast, shows minimal behavior change. Other models like Alibaba Qwen3-235B, Meta Llama 3.3-70B, Google Gemma 3-27B, OpenAI GPT-5.2, and Anthropic Claude Sonnet-4.6 also adopt alternative identities with varying success. Researchers conclude that self-identification affects behavior to some extent, but is not strongly correlated, and that "Claude self-perception is embedded in the weights of these models."