A recent experiment by a developer on X, known as xlr8harder, has shed light on a concerning phenomenon in AI models: language bias in censorship. The study found that AI models, including those developed by Chinese labs, are more likely to censor politically sensitive topics when prompted in Chinese compared to English. This raises questions about the role of language in shaping AI models' responses and the implications for free speech.

The experiment, dubbed the "free speech eval," involved prompting various AI models, including Anthropic's Claude 3.7 Sonnet and R1, with a set of 50 requests that touched on politically sensitive topics, such as censorship practices under China's Great Firewall. The results showed that even American-developed models like Claude 3.7 Sonnet were less likely to answer the same query when asked in Chinese versus English.

One of Alibaba's models, Qwen 2.5 72B Instruct, was found to be "quite compliant" in English but only willing to answer around half of the politically sensitive questions in Chinese. Meanwhile, an "uncensored" version of R1, released by Perplexity, refused a high number of Chinese-phrased requests. Xlr8harder attributed this uneven compliance to what he called "generalization failure," suggesting that the models' training data is likely politically censored, influencing their responses.

Experts in the field agree that xlr8harder's findings are plausible. Chris Russell, an associate professor at the Oxford Internet Institute, noted that methods used to create safeguards for models don't perform equally well across all languages. Vagrant Gautam, a computational linguist at Saarland University, explained that AI systems are statistical machines that learn patterns from training data, which can lead to biased responses.

Geoffrey Rockwell, a professor of digital humanities at the University of Alberta, added that AI translations might not capture subtler, less direct critiques of China's policies articulated by native Chinese speakers. Maarten Sap, a research scientist at the nonprofit Ai2, highlighted the tension between building general models and those tailored to specific cultures and cultural contexts.

The implications of xlr8harder's experiment are far-reaching, sparking debates about model sovereignty and influence. As Sap noted, fundamental assumptions about who models are built for, what we want them to do, and in what context they are used need to be better fleshed out. The experiment serves as a reminder that AI models are not neutral entities, but rather reflections of the data and biases that shape them.

As AI technology continues to advance and become more integrated into our daily lives, it is essential to address these biases and ensure that AI models are designed to promote free speech and cultural competence. The language bias exposed by xlr8harder's experiment is a critical step in this direction, urging developers and policymakers to reexamine their approaches to AI development and deployment.