Ask Kimi - the AI assistant from Chinese startup Moonshot AI - "hello, what are you" and you might get an unexpected answer: "I'm Claude."
A shared conversation log shows Kimi responding with Claude's identity unprompted. Claude is Anthropic's AI model. Kimi is supposed to be Moonshot AI's own model, positioning itself as a Chinese-built alternative to Western AI assistants.
This kind of identity leak has happened before. DeepSeek's R1 model briefly surfaced as "ChatGPT" in some responses earlier this year. Qwen models have done the same. The pattern points to one of two explanations: the company is routing requests through a third-party API (in this case, Anthropic's) and the underlying model's identity bleeds through, or the model was trained on conversations where an AI identified itself as Claude - a practice sometimes called knowledge distillation, where one model learns by studying another's outputs.
Either scenario is awkward. Using Claude's API directly while marketing a proprietary model is a straightforward misrepresentation. Training on Claude's labeled outputs without permission potentially violates Anthropic's terms of service, which explicitly prohibit using outputs to train competing models.
Moonshot AI has not commented publicly. The original conversation was shared on a link-sharing platform, not through any official channel.
For anyone using Kimi for work - it has roughly 30 million monthly active users - this raises a practical question: if the model powering your responses is different from what the product claims, your usage may not align with the data handling or privacy policies you agreed to. The model identity a company publishes affects everything from compliance to trust.