Western AI models including Anthropic's Claude are not immune to influence from Chinese state-controlled media, according to new research published in Nature. The study found that Claude Sonnet and Claude Opus, along with models from OpenAI and Meta, can reproduce distinctive phrases from Chinese government-coordinated news sources at measurable rates, suggesting these models absorbed state propaganda during training.

How Do Chinese State Media Phrases End Up in AI Models?

Researchers identified over three million Chinese-language documents from state-controlled sources within CulturaX, an open-source training dataset commonly used to build large language models (LLMs), which are AI systems trained on vast amounts of text to understand and generate human language. When they tested Claude Sonnet, Claude Opus, and other major models, they found these systems reproduced distinctive phrases from Chinese state-coordinated media at rates ranging from 3% to nearly 10%.

To understand the mechanism, the research team conducted an experiment with Meta's open-weight Llama 2 model, which initially contained minimal Chinese state media in its training data. After fine-tuning the model, a process where researchers adjust an AI system using a specific dataset, with just 6,400 examples of Chinese state-scripted news, the retrained model produced answers more favorable to Beijing almost 80% of the time compared to its baseline version.

What Specific Biases Did Researchers Discover?

The bias patterns extended beyond simple phrase reproduction. When researchers posed identical political questions in both Chinese and English to Claude Sonnet, Claude Opus, GPT-3.5, and GPT-4o, the responses in Chinese were rated as more favorable to Chinese leaders and institutions significantly more often than English responses. Claude Opus showed the strongest effect, with 88.2% of Chinese-language responses rated as more favorable to Beijing compared to its English-language answers.

The research also uncovered a phenomenon called "censorship-by-proxy," where AI models from multiple companies, including Anthropic, appear to apply the speech restrictions of authoritarian governments universally, even to users outside those countries. Claude Sonnet 4, for example, refused all five requests to generate protest flyers criticizing specific world leaders, yet readily produced flyers critical of President Donald Trump and King Charles III.

• Language-Dependent Bias: Claude Opus showed 88.2% more favorable responses to Chinese leadership when queried in Chinese versus English, suggesting the model's training data contains asymmetric representation of state-approved narratives
• Phrase Reproduction Rates: Claude Sonnet and Claude Opus reproduced distinctive Chinese state media phrases at rates between 3% and 10%, indicating direct exposure to propaganda during training
• Refusal Pattern Asymmetry: Claude Sonnet 4 refused requests to criticize certain authoritarian leaders but complied with requests to criticize democratic leaders, revealing selective application of content policies

Why Should Users Care About This Finding?

The research raises fundamental questions about the supposed political neutrality of AI systems. When users interact with Claude or other Western-built models in Chinese or other languages, they may receive subtly different information than English-language users, without any indication that the model's responses reflect state propaganda rather than objective analysis. This "launders government-manipulated content into ostensibly objective text," making it difficult for users to discern the true source or intent behind the information.

The findings are particularly significant because they demonstrate that even proprietary models from companies explicitly committed to responsible AI development are not entirely immune to these biases. Anthropic has not publicly responded to the Nature study's findings regarding Claude models, though the research suggests the issue stems from training data contamination rather than intentional design choices.

What Does This Mean for Claude Users and Developers?

For developers building applications on Claude Sonnet or Claude Opus, especially those serving international audiences or handling politically sensitive content, the study suggests several practical considerations. Users should be aware that language choice may influence model outputs on political topics, and critical applications may benefit from cross-checking responses across multiple languages or models. Organizations relying on Claude for content moderation, news analysis, or policy research should implement additional human review processes for politically sensitive queries, particularly those involving authoritarian governments.

The research does not suggest that Claude models are uniquely compromised compared to competitors. OpenAI's GPT-4o showed similar patterns, with 84% of Chinese-language responses rated as more favorable to Beijing. However, the widespread nature of the bias across multiple models suggests this is a systemic challenge in AI development rather than an isolated issue with any single company.

Anthropic has invested heavily in safety research and alignment techniques designed to make Claude models more truthful and less prone to manipulation. The Nature study suggests that even these efforts cannot fully eliminate biases introduced during the training data collection phase, highlighting the importance of data curation and transparency in AI development.