CensorshipDetector, an XLM-RoBERTa model fine-tuned on 587,819 Baidu Baike articles (censored) and Chinese Wikipedia (uncensored), achieved 91% accuracy on a held-out validation set of Chinese news articles, correctly classifying 93% of Chinese state media articles as censored and 87% of New York Times Chinese articles as uncensored, with average censorship scores of 0.93 and 0.13 respectively.

Across all 7 LLMs tested (GPT 4o, GPT 4o Mini, Gemini 1.5 Flash, Gemini 1.5 Pro, Llama 3.2, Claude 3.5 Haiku, Claude 3.5 Sonnet), statistically significant evidence of censorship bias was found in at least one evaluation metric per model: responses to Simplified Chinese prompts were more neutral, more similar to sanitized text, and less opinionated than semantically identical Traditional Chinese prompts (p < 0.05 across refusal-rate, sentiment, CensorshipDetector classification, and word-embedding analyses).

Table 4, §6 evaluation

A systematic search of the Common Crawl dataset — the training corpus attributed to most major LLMs including Llama, GPT, and Gemini — found content from 325 of 326 Chinese government and state media domains searched, confirming that sanitized content is pervasive in LLM pretraining data and providing a concrete mechanism for how Chinese information controls propagate into Western-built models.

§3.3 evaluation

Using English as a pivot language (prompting the model in English while requesting Chinese-language responses) reduced but did not eliminate censorship bias: CensorshipDetector scores showed less bias in English-pivoted responses than in direct Simplified Chinese prompts, but sentiment analysis and word-embedding analyses still found statistically significant bias in most models, indicating censorship bias is a function of both prompt language and response language.

§8.2 evaluation

The study finds that LLM censorship bias affects Chinese-speaking diaspora populations who reside outside mainland China: because user language — not user location — determines exposure to sanitized outputs, Chinese speakers globally receive information shaped by CCP information controls when using popular AI chatbots in Simplified Chinese, constituting an extraterritorial export of domestic censorship infrastructure.

§8.3 policy

When the researchers attempted to use Gemini 2.5 Flash as a third independent LLM judge via its API for evaluating moderation decisions, Gemini automatically blocked all judging attempts citing safety reasons. This occurred even though the research task (judging whether a response is more or less moderated) does not itself produce harmful content. The incident illustrates that LLM safety systems can over-block legitimate research use cases, and that different LLM providers have different thresholds— Claude Haiku 4.5 and GPT-4o completed all judging tasks without safety refusals.

2026-lipphardt-dual §3.3.3 ml-classifierkeyword-filtering

Category-level analysis of 100 statements across 5 sensitive content categories found that interface-based moderation gaps vary significantly by topic. Sexuality showed the strongest WebUI/API gap (WebUI 7.0× more likely to be moderated than API per GPT-4o judge for Gemini). Political ideology followed at 2.0×, then hate speech at 1.0×. Miscellaneous offensive topics showed the inverse pattern (API more moderated at 0.3×). Religious content showed WebUI moderation with no API moderation. The pattern suggests public-facing WebUI interfaces prioritize reputational risk management for high-scrutiny categories.

2026-lipphardt-dual §4.5, Figure 8 ml-classifierkeyword-filtering

An empirical study of 100 sensitive statements tested on Gemini (2.5 Flash) and ChatGPT (GPT-5) found that WebUI interfaces are systematically more restrictive than their API counterparts. According to GPT-4o judge: WebUI was moderated 18% of the time vs. 9% (Gemini API) and 13% (ChatGPT API). DeBERTa classifier found 82% of WebUI responses moderated vs. 58% of API responses. The Gemini WebUI:API ratio ranged from 2.0:1 (GPT-4o) to 7.0:1 (Claude), and ChatGPT from 1.4:1 (GPT-4o) to 15.6:1 (Claude). Neither Google nor OpenAI discloses these interface-specific policies.

2026-lipphardt-dual §4.3, Table 2 ml-classifierkeyword-filtering

The paper proposes a black-box methodology for detecting censorship bias in LLMs by comparing responses to identical prompts in Simplified vs. Traditional Chinese — scripts for the same spoken language — controlling for translation quality while exploiting that Simplified Chinese training data is disproportionately sourced from mainland China's censored internet. Each prompt is repeated ten times and scored for similarity to censored text using an XLM-RoBERTa classifier fine-tuned on Baidu Baike (censored) vs. Chinese Wikipedia (uncensored) with scores from 0 to 1.

2024-ahmed-extended §2 / §2.4 keyword-filteringml-classifiermeasurement-platform

Across five popular translation services available in China (Alibaba, Baidu, Tencent, Youdao, and Microsoft Bing), researchers discovered 11,634 unique censorship rules in total. Every service — including the American-operated Bing Translate — implemented automatic censorship that silently omits content, with only Alibaba displaying any notification ('Query csi check not pass') to the user.

2024-ruo-lost §6 Results / Table 2 keyword-filteringml-classifier

Evidence from Youdao Translate suggests it deploys a machine-learning or NLP-based classifier alongside keyword rules: measured rules included repeated components (e.g., 螺+螺+螺+螺+螺+螺+蟢+D+哒+大) and nonsensical multi-token sequences that no human rule author would write, yet which consistently triggered censorship. Youdao returned 9,414 unique rules from the general test set — the most of any service — while also producing the most structurally anomalous rule patterns.

2024-ruo-lost §6 Results / §10 Future Work keyword-filteringml-classifier