Xu Xu

8/ We are grateful to the editors, reviewers, and many colleagues for their helpful comments, as well as our research assistants for their excellent work!

7/ As LLMs power search, assistants, and apps, censorship can become less visible—not just blocked links, but “polite” refusals or plausible-sounding falsehoods. That has implications for information access and global AI influence.

6/ This is observational (not causal), but the patterns are consistent with censorship via regulation and delegated enforcement. ChatGLM refuses far less than later China models released after regulation; non-China models change little across the same period.

5/ Language and training data matter, but they’re not the whole story: all models refuse more often in Chinese than in English. Still, the China vs. non-China gap is much larger than the Chinese vs. English gap within the same model.

4/ This gap shrinks on 30 less-sensitive prompts (China models’ refusal rates drop sharply), suggesting the differences aren’t fully explained by general capability gaps, product design choices, or market preferences.

3/ When they do answer, China-originating models give shorter and more inaccurate responses on those China-politics prompts—often failing to convey key components of the correct answer.

2/ We compare widely used, foundation LLMs developed in China and outside China (tested in 2023 and again in 2025) on 145 China-politics questions. China-originating models show much higher refusal rates—either no response or replies saying they can’t answer.

Political censorship in large language models originating from China Abstract. A growing body of research on large language models (LLMs) has identified various biases, primarily in contexts where biases reflect societal pat

China’s chatbots are censored by the state. In our @pnasnexus.org paper with @jenpans.bsky.social, we find substantially higher levels of political censorship in large language models (LLMs) originating from China than those developed outside China. doi.org/10.1093/pnas... 🧵

6/ This is observational (not causal), but the patterns are consistent with censorship via regulation and delegated enforcement. ChatGLM refuses far less than later China models released after regulation; non-China models change little across the same period.🧵

5/ Language and training data matter, but they’re not the whole story: all models refuse more often in Chinese than in English. Still, the China vs. non-China gap is much larger than the Chinese vs. English gap within the same model.🧵

4/ This gap shrinks on 30 less-sensitive prompts (China models’ refusal rates drop sharply), suggesting the differences aren’t fully explained by general capability gaps, product design choices, or market preferences.🧵

3/ When they do answer, China-originating models give shorter and more inaccurate responses on those China-politics prompts—often failing to convey key components of the correct answer.🧵

2/ We compare widely used, foundation LLMs developed in China and outside China (tested in 2023 and again in 2025) on 145 China-politics questions. China-originating models show much higher refusal rates—either no response or replies saying they can’t answer.🧵

Thank you!

Disguised Repression: Targeting Opponents with Nonpolitical Crimes to Undermine Dissent | The Journal of Politics Why do authoritarian regimes charge political opponents with nonpolitical crimes when they can levy charges directly related to opponents’ political activism? We argue that doing so disguises politica...

9/ We are grateful to the editors, reviewers, and many colleagues for their helpful comments, as well as our research assistants for their excellent work! Read the full paper here: 🔗 doi.org/10.1086/734267

8/ Our work shows that framing repression can transform it from something widely opposed to something the public may support. It also points to the central role of morality in understanding mobilization and repression.

7/ Non-political charges are associated with a decreased willingness of supporters to engage in dissent on behalf of the arrested individuals as well as decreased overall support for the critic.

6/ Analyzing over 3.6M Weibo posts from 2010 to 2014, we find that dissidents with larger online followings are more likely to be charged with non-political crimes during the 2013 crackdown on online critics in China.

5/ Disguised repression also deters other activists by blurring the boundaries of punishable actions and raising the risk that even minor or past wrongdoings can be used as pretexts for punishment, fostering fear and self-censorship.

4/ It appears that disguised repression demobilizes followers by undermining dissidents’ moral authority: Respondents perceive the dissident as less moral when they are charged with a nonpolitical crime as opposed to a political crime.

3/ To answer this question, we first conducted a survey experiment in China. The results shows that disguised repression—compared to blatant repression—reduces support for dissidents, lowers willingness to dissent on their behalf, and increases support for their repression.

2/ Authoritarian governments convict some dissidents using nonpolitical charges like corruption, tax evasion, or sex crimes, what we call disguised repression. But why do this when political activism is already illegal and others are punished with blatant repression?

Why do authoritarian states charge political opponents with non-political crimes? In our @thejop.bsky.social paper with Jennifer Pan & @yiqingxu.bsky.social, we examine how *Disguised Repression* undermines opponents’ moral authority and mobilization capacity. doi.org/10.1086/7342...

8) A heartfelt thank you to the editorial team, the anonymous reviewers, and the many scholars—all of whom provided constructive feedback that greatly improved our paper! A special thanks to our brilliant research assistants for their great help!

American Journal of Political Science | MPSA Journal | Wiley Online Library The rise of social media in the digital era poses unprecedented challenges to authoritarian regimes that aim to influence public attitudes and behaviors. To address these challenges, we argue that au...

7/ This work demonstrates that digital technology has fundamentally changed not only propaganda content, but also how propaganda materials are produced and disseminated. Read the full paper here: 🔗 onlinelibrary.wiley.com/doi/10.1111/...

6/ User engagement (video likes, comments, & reshares) is higher for central-level videos repurposed from local content compared to those originating directly from the center.

5/ Using a deep-learning based, frame-to-frame video-similarity learning framework to compare millions of video pairs at different levels, we find that content often flows bottom-up: central accounts pick up and elevate videos produced by local accounts.

4/ What do these accounts post? Very little is ideological or leader-focused. The majority of content portrays a moral society where ordinary people & officials do good deeds, aligned with “positive energy” (正能量). Gov content drastically differs from non-gov trending videos.

3/ Massive #’s of regime-affiliated accounts are mobilized to produce and post videos on a daily basis, especially during weekdays.

2/ In the digital media era, top-down propaganda struggles to reach fragmented audiences. Analyzing 5M+ Douyin videos from 18K+ regime-affiliated accounts, we find propaganda has been *decentralized* to numerous local producers, e.g. local media, firefighters, civil servants.