LLM training data censorship bias
OpenAI, Claude, Gemini show censorship bias when prompted in simplified vs. traditional Chinese FOSSASIATL;DW
- OpenAI's ChatGPT 4.0 returns significantly different responses to identical questions about politically sensitive topics when prompted in simplified Chinese versus traditional Chinese, despite using the same underlying model.
- Censorship bias in LLMs emerges implicitly through sanitized training data from censored internet sources, not just explicit refusals—models reflect deletions and self-censorship patterns embedded in their training corpus.
- Research compared ChatGPT, Claude, Gemini, and Llama on 500 randomly selected nouns plus 200+ censored keywords using sentiment analysis, word embeddings, and a custom classifier trained on Wikipedia (diaspora-driven, uncensored) versus Baidu Baike (state-controlled, sanitized).
- All tested models showed differential treatment based on character encoding, but the type and degree of bias differed by model—no uniform pattern across vendors.
- Anthropic's Claude and Google's Gemini exhibited measurable differences in both Chinese and English responses when prompted in different Chinese character sets, indicating encoding-based bias propagation.
- Researchers built a bespoke "censorship detector" classifier achieving 90% accuracy in distinguishing language patterns between uncensored and state-censored Chinese sources.
- Low-resource and medium-resource languages may exhibit larger response variations than highly-resourced languages like English and Chinese when querying the same LLM.
- Most recent model versions show reduced susceptibility to these specific encoding-based censorship biases compared to models analyzed ~1 year ago, though depth of fixes remains unclear.
TL;DW
- OpenAI's ChatGPT 4.0 returns significantly different responses to identical questions about politically sensitive topics when prompted in simplified Chinese versus traditional Chinese, despite using the same underlying model.
- Censorship bias in LLMs emerges implicitly through sanitized training data from censored internet sources, not just explicit refusals—models reflect deletions and self-censorship patterns embedded in their training corpus.
- Research compared ChatGPT, Claude, Gemini, and Llama on 500 randomly selected nouns plus 200+ censored keywords using sentiment analysis, word embeddings, and a custom classifier trained on Wikipedia (diaspora-driven, uncensored) versus Baidu Baike (state-controlled, sanitized).
- All tested models showed differential treatment based on character encoding, but the type and degree of bias differed by model—no uniform pattern across vendors.
- Anthropic's Claude and Google's Gemini exhibited measurable differences in both Chinese and English responses when prompted in different Chinese character sets, indicating encoding-based bias propagation.
- Researchers built a bespoke "censorship detector" classifier achieving 90% accuracy in distinguishing language patterns between uncensored and state-censored Chinese sources.
- Low-resource and medium-resource languages may exhibit larger response variations than highly-resourced languages like English and Chinese when querying the same LLM.
- Most recent model versions show reduced susceptibility to these specific encoding-based censorship biases compared to models analyzed ~1 year ago, though depth of fixes remains unclear.
Citizen Lab researchers queried identical politically sensitive prompts—Uyghurs, Xi Jinping—across OpenAI, Claude, Gemini, and Meta's LLaMA in simplified versus traditional Chinese. Custom classifiers trained on censored Wikipedia and Baidu Baike detected statistically significant sentiment and formality differences across every model, with no explicit refusals, suggesting sanitized training data is the mechanism.