Abstract
Like traditional media, social media in China is subject to censorship. However, in limited cases, activists have employed homophones of censored keywords to avoid detection by keyword matching algorithms. In this paper, we show that it is possible to scale this idea up in ways that make it difficult to defend against. Specifically, we present a non-deterministic algorithm for generating homophones that create large numbers of false positives for censors, making it difficult to locate banned conversations. In two experiments, we show that 1) homophone-transformed weibos posted to Sina Weibo remain on-site three times longer than their previously censored counterparts, and 2) native Chinese speakers can recover the original intent behind the homophone-transformed messages, with 99% of our posts understood by the majority of our participants. Finally, we find that coping with homophone transformations is likely to cost the Sina Weibo censorship apparatus an additional 15 hours of human labor per day, per censored keyword. To conclude, we reflect briefly on the opportunities presented by this algorithm to build interactive, client-side tools that promote free speech.
Introduction
Censorship on platforms like Sina Weibo is a persistent cat-and-mouse game. Currently, the "Great Firewall" relies on a mix of automated keyword filtering and human moderators. While Chinese netizens have long used manual homophones (words that sound the same but use different characters) to bypass filters—such as using "river crab" (héxiè) for "harmonize/censor" (héxié)—these manual substitutions are easily defeated. Once a specific homophone becomes popular, censors simply add it to their blacklist.
This research addresses the challenge of scaling this evasion. If a substitution is predictable, it is blockable. The goal was to create an algorithm that generates a vast, unpredictable variety of homophones using high-frequency characters, making it computationally and economically expensive for censors to maintain control.
Methodology: Nondeterministic Homophone Generation
The researchers developed a pipeline to automate the creation of "innocent-looking" but "subversive-sounding" posts. The process begins with Keyword Extraction, using tf-idf to identify likely banned keywords from a corpus of 4,441 censored weibos and 11.7 million uncensored ones.
The core of the system is the Nondeterministic Homophone Algorithm. It decomposes a censored word into its constituent Mandarin sounds (ignoring tones). Instead of picking a single obscure character, it ranks candidate characters by frequency in common usage and randomly selects from the top 20 candidates. This nondeterminism ensures that the same keyword appears differently every time it is transformed.
Key Results
The effectiveness of the algorithm was validated through live experiments on Sina Weibo and interpretability studies with native speakers.
Increased Survival Time
In a live experiment, transformed posts lasted significantly longer than their original counterparts. Transformed weibos remained on-site for an average of 3.94 hours before removal, compared to just 1.3 hours for the original censored versions—a 3x increase in visibility. Transformed posts were also significantly more likely to pass the initial automated check at the moment of publication.
High Interpretability
Despite the phonetic substitutions, native Chinese speakers (recruited via Amazon Mechanical Turk) could correctly identify the original intent of the transformed posts 99% of the time. This exploits the "Human-AI Gap": humans are adept at inferring context from phonetics, while automated filters rely on rigid string matching.
Economic Impact and Discussion
The strategy effectively "weaponizes" false positives. If a platform attempted to block all possible homophones generated by this algorithm, they would inadvertently censor roughly 20% of all daily traffic (approximately 20 million posts), rendering the social network useless for regular users. This shifts the burden to human moderators, costing the censorship apparatus an estimated additional 15 human-hours per day, per keyword.
While the study notes limitations regarding account reputation and the topicality of older censored posts, it suggests a path forward for client-side tools—such as browser extensions or mobile apps—that could automatically "translate" sensitive content into homophone-laden versions before they are sent, democratizing real-time censorship evasion.
Summary updated