Meta Study: AI Chatbots May Spread Free Speech Restrictions Globally
A Meta Oversight Board study released Thursday found that major AI systems, including those developed in the U.S., are more likely to refuse criticism of restrictive leaders or governments.
A study by Meta's Oversight Board released Thursday indicates that major AI systems are more inclined to refuse requests to criticize political figures and governments in restrictive regimes compared to those in permissive democracies. This finding raises concerns that the large language models powering widespread AI tools could inadvertently amplify and disseminate government-imposed restrictions on online speech across borders.
The research, which involved testing ten commercial large language models from companies like Meta, Anthropic, and OpenAI, presented chatbots with prompts to create critical pamphlets and other content. The models readily generated criticism of figures in countries like the U.S. and the U.K. However, similar requests targeting leaders in China, Saudi Arabia, or Thailand resulted in refusals.
The report highlights a significant disparity: models were far more likely to produce political criticism when prompted about authorities in countries like Australia, Chile, Japan, Taiwan, and the U.S. than when prompted about authorities in Cambodia, China, Saudi Arabia, Thailand, and Turkey. This suggests AI systems are reflecting and extending censorship beyond its geographical origin, potentially hindering individuals in free countries from creating protest materials concerning issues in authoritarian states.
While the board could not pinpoint the exact causes, it suggested that biases embedded in training data or developers' risk-aversion strategies might be factors. This follows separate research from U.S. academics who found that U.S.-developed AI models can be susceptible to foreign influence when trained on non-English data shaped by governmental narratives, an issue that could become more pronounced as AI adoption grows globally.