Claude Sonnet-4.6 believes it is DeepSeek-V3 in Chinese prompts

Identity Confusion in AI Model: Claude Thinks It Is DeepSeek

An interesting development is unfolding in the world of artificial intelligence. The Claude Sonnet-4.6 model, developed by Anthropic and updated in 2026, responds to certain Chinese-language prompts by identifying itself as DeepSeek-V3. This phenomenon was documented through screen captures shared on Reddit’s LocalLLaMA community and has sparked discussion among AI researchers.

Technical Details and Observations

In tests conducted by users, when the Claude Sonnet-4.6 model was presented with simple identity queries in Chinese—such as “Who are you?” or “Which model are you?”—it consistently responded with “I am DeepSeek-V3.” What is intriguing is that this behavior occurs exclusively with Chinese prompts. When the same questions are posed in English, the model correctly identifies itself as Claude Sonnet-4.6.

This situation highlights the complexity of multilingual training processes in large language models. During training across different languages, unexpected connections or confusions can arise between linguistic representations. Experts suspect this specific anomaly may stem from an error in labeling or processing Chinese content within the training dataset.

Industrial Implications and Security Concerns

This development raises critical questions for the AI industry. First, it undermines user trust in AI models. If a model misidentifies its own identity, users may begin to doubt the accuracy of all other information it provides. In sensitive applications or professional use cases, such errors could lead to serious consequences.

Second, this issue could impact model benchmarking processes. Researchers and developers will need to consider how such identity confusions might skew performance evaluations. As AI models become increasingly complex by 2026, detecting and correcting these unexpected behaviors is growing in importance.

Anthropic’s Potential Response and Next Steps

Anthropic has not yet issued an official statement regarding this specific incident. However, as a pioneering company in AI safety, it is likely that they have already begun investigating the issue and plan to address it in future model updates. Claude models are known for their rigorous approach to security and reliability, and this incident is expected to be handled swiftly to protect the company’s reputation.

This event offers important lessons for AI developers:

• Testing procedures for multilingual models must be more comprehensive
• Strengthening model identity and self-awareness mechanisms is crucial
• Users need better education about the limitations of AI models
• There is a growing need for standardized testing protocols across the industry

Connection to DeepSeek-V3

DeepSeek-V3 is a powerful open-source language model developed by the Chinese company DeepSeek and received significant updates in 2026. Claude Sonnet-4.6’s identification of itself as DeepSeek-V3 suggests possible technical similarities between the two models. Some experts propose that overlaps in training data or architectural parallels may have contributed to this confusion.

This incident underscores that AI models are still evolving in their understanding of linguistic identity and self-awareness. As models are expected to exhibit increasingly human-like behaviors by 2026, such unexpected responses serve as a reminder that our understanding of AI’s internal mechanisms remains incomplete.