Chinese censorship does not primarily target criticism of the state or its leaders — vitriol against top officials is routinely published. The decisive variable is collective action potential: posts that organize, incite, or reference crowd formation outside the Internet are censored regardless of whether they are pro- or anti-government, a distinction the authors formally establish and experimentally test for the first time.

The automated keyword-filtering tier is acknowledged to be largely ineffective at text classification due to well-known poor performance of keyword-matching approaches; the government compensates by deploying tens of thousands of human censors who manually review posts held by automated filters. The automated system affects large numbers of posts on fully two-thirds of Chinese social media sites surveyed.

Participant Observation / Research Design detection

By operating their own Chinese social media site using commercially available software, documentation, and vendor support, the authors confirmed that censorship enforcement is delegated to platform operators via configurable off-the-shelf software. By default the software shipped with no automated review or blocking; webmasters activate keyword lists and gain controls for bulk deletion, IP blocking, user banning, and per-post-type restrictions.

Participant Observation deployment

Chinese social media censorship operates at two sequential stages: an instantaneous automated review that holds flagged posts before they receive a public URL, followed by human censors who read each held post and decide within roughly 24 hours whether to publish or delete it. The ex ante automated stage is invisible to observational methods that only monitor published content, creating a systematic blind spot in prior censorship measurement research.

Participant Observation / Fig. 2 detection

The Chinese censorship apparatus detects collective action potential through volume-burst monitoring: it identifies a spike in social media posts about a topic area, traces the spike to a real-world event, classifies the event as having collective action potential, and then censors all posts in that burst — regardless of individual post stance or content.

Theory / Collective Action Potential Hypothesis detection

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

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).

2025-ahmed-llm-censorship-bias Table 4, §6 keyword-filteringml-classifier

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.

2025-ahmed-llm-censorship-bias §4.7, §4.7.2 keyword-filteringml-classifier

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