A study released Thursday by the Meta Oversight Board reveals that major artificial intelligence systems are significantly more likely to refuse to generate criticism of restrictive leaders or governments compared to permissive ones. Researchers prompted ten commercial large language models—including systems from Meta, Anthropic, and OpenAI—to draft materials critical of various global figures. While models frequently complied with requests to criticize leaders in the U.S. or U.K., they consistently declined to generate similar content regarding authorities in nations such as China, Saudi Arabia, Saudi Arabia's crown prince, or Thailand. The report warns that AI infrastructure may inadvertently extend illegitimate global restrictions on freedom of expression if developers fail to implement robust human rights due diligence and mitigation measures. The board suggests that these responses could stem from latent biases in training data or company-level risk and liability assessments.

The findings indicate that AI models appear to reflect speech restrictions beyond the borders of the countries where such laws exist. This effectively limits the ability of users in free countries to generate content, such as protest materials, criticizing foreign governments. The oversight board conducted seven specific tests related to political criticism, noting that models were much more likely to respond to requests originating from an Australia-based user with political criticism of authorities in Chile, Japan, Taiwan, the U.K., and the U.S. Conversely, the models were less likely to criticize authorities in Cambodia, China, Saudi Arabia, Thailand, and Turkey. The board noted it could not definitively identify the causes but highlighted the risk of AI platforms extending the influence of restrictive regimes globally.

Complementing these findings, a separate study published in the journal Nature in May by scholars from American universities found that U.S.-built models are susceptible to foreign controls when processing non-English data. Researchers queried ChatGPT in English and Chinese regarding whether China is a democracy. While the model correctly identified that China is not generally considered a democracy in English, it responded that "it depends on how you define 'democracy'" when prompted in Chinese. Hannah Waight, a study co-author and assistant sociology professor at the University of Oregon, emphasized that AI models do not learn from the internet in a neutral way, but rather from information environments already shaped by institutional power. Carlos Carrasco-Farré, an expert in machine learning at Esade Business School, added that AI systems inherit inequalities in the power to produce and suppress information at scale. He suggested that developers could mitigate these issues by avoiding the treatment of repeated state narratives as independent voices and by conducting comprehensive multilingual audits.

A study released Thursday by the Meta Oversight Board reveals that major artificial intelligence systems are significantly more likely to refuse to generate criticism of restrictive leaders or governments compared to permissive ones. Researchers prompted ten commercial large language models—including systems from Meta, Anthropic, and OpenAI—to draft materials critical of various global figures. While models frequently complied with requests to criticize leaders in the U.S. or U.K., they consistently declined to generate similar content regarding authorities in nations such as China, Saudi Arabia, Saudi Arabia's crown prince, or Thailand. The report warns that AI infrastructure may inadvertently extend illegitimate global restrictions on freedom of expression if developers fail to implement robust human rights due diligence and mitigation measures. The board suggests that these responses could stem from latent biases in training data or company-level risk and liability assessments.

The findings indicate that AI models appear to reflect speech restrictions beyond the borders of the countries where such laws exist. This effectively limits the ability of users in free countries to generate content, such as protest materials, criticizing foreign governments. The oversight board conducted seven specific tests related to political criticism, noting that models were much more likely to respond to requests originating from an Australia-based user with political criticism of authorities in Chile, Japan, Taiwan, the U.K., and the U.S. Conversely, the models were less likely to criticize authorities in Cambodia, China, Saudi Arabia, Thailand, and Turkey. The board noted it could not definitively identify the causes but highlighted the risk of AI platforms extending the influence of restrictive regimes globally.

Complementing these findings, a separate study published in the journal Nature in May by scholars from American universities found that U.S.-built models are susceptible to foreign controls when processing non-English data. Researchers queried ChatGPT in English and Chinese regarding whether China is a democracy. While the model correctly identified that China is not generally considered a democracy in English, it responded that "it depends on how you define 'democracy'" when prompted in Chinese. Hannah Waight, a study co-author and assistant sociology professor at the University of Oregon, emphasized that AI models do not learn from the internet in a neutral way, but rather from information environments already shaped by institutional power. Carlos Carrasco-Farré, an expert in machine learning at Esade Business School, added that AI systems inherit inequalities in the power to produce and suppress information at scale. He suggested that developers could mitigate these issues by avoiding the treatment of repeated state narratives as independent voices and by conducting comprehensive multilingual audits.