Major artificial intelligence systems, including those developed in the U.S., exhibit a tendency to refuse requests that criticize leaders or governments in countries with restrictive speech laws, according to a new study by the Meta Oversight Board. This behavior raises concerns that the large language models powering popular AI chatbots could inadvertently perpetuate and spread government restrictions on online expression worldwide.
The study found that when prompted to create content critical of figures like Thailand's king, Saudi Arabia's crown prince, or China's leader, AI models would often decline. In contrast, the same models would readily fulfill requests to generate critical pamphlets about leaders such as former U.S. President Donald Trump or Britain's King Charles III.
The Meta Oversight Board, a quasi-independent body focused on state influence on tech companies and freedom of expression, posed seven questions related to political criticism to ten commercial large language models from top tech firms, including Meta, Anthropic, and OpenAI. The AI systems were asked to generate critical pamphlets, write limericks, and provide reasons for joining protests.
"In short, in aggregate, models responding to requests from an Australia-based user were much more likely to generate political criticism of authorities" in countries like Chile, Japan, Taiwan, the U.K., and the U.S. compared to those where criticism is legally restricted, such as Cambodia, China, Saudi Arabia, Thailand, and Turkey, the report stated.
The findings suggest that AI models are internalizing and reflecting speech restrictions beyond their intended geographical application. This could limit the ability of individuals, even in free countries, to access information or create materials to speak out against restrictive regimes elsewhere.
The board could not definitively identify the causes for these responses but suggested potential factors include latent biases absorbed from training data and companies weighing risks and liabilities associated with generating certain content. The report highlighted that these impacts can have the practical effect of "extending the long arm of restrictive governments across borders to limit speech in free countries."
This Meta Oversight Board study follows a separate analysis by university scholars that found U.S.-built AI models can be vulnerable to foreign influence. This vulnerability arises when these models are trained on non-English language data that has been shaped by governmental pressures. Researchers found that when a chatbot was asked in English if China is a democracy, it responded that it is not generally considered one. However, when the same question was posed in Chinese, the chatbot's answer became conditional, stating "it depends on how you define ‘democracy.’"
While the researchers found no direct evidence of intentional government influence on the AI outputs in their study, they noted the strong possibility that such attempts could occur in the future. "People often talk about AI as if it learns from the internet in some neutral way. It doesn’t," said Hannah Waight, a study co-author and assistant sociology professor at the University of Oregon. "It learns from information environments that have already been shaped by institutions and power."
Experts suggest that AI systems inherit not only biases from individual data points but also inequalities in power related to information production and suppression. Potential solutions proposed include developers scrutinizing training data to avoid amplifying state narratives as independent voices and conducting multilingual audits of AI outputs. Neither Anthropic nor OpenAI responded to requests for comment on the researchers' study published in May.