The Meta Oversight Board recently released a study uncovering critical insights about AI systems. The research shows that AI-powered chatbots, including those developed by U.S.-based companies, often decline to criticize certain leaders from restrictive governments. This has sparked concerns that AI could be unintentionally spreading government influence over online speech.
The report highlights the risk these AI systems pose to freedom of expression worldwide. If developers do not conduct thorough human rights evaluations, they could inadvertently build AI frameworks that impose illegitimate speech restrictions globally. This statement underscores the gravity of the findings, which stem from studies involving top tech firms, including Meta, Anthropic, and OpenAI.
The oversight board posed seven questions to the AI models, focusing on political criticism across different governance types. The responses indicated that chatbots based in relatively free countries attracted far more political criticism compared to those from places where such actions are penalized, like China and Saudi Arabia.
The report suggests that AI models might reflect speech restrictions beyond their development countries. For instance, a user in Australia might struggle to generate protest materials against actions by authorities in China due to the biases embedded in AI systems. These impacts effectively extend the restrictive control of some governments across international borders.
Causes for the differential responses in the AI models remain undetermined. Potential factors include latent biases in training data and companies weighing risks and liabilities. The Meta Oversight report coincides with a separate academic study showing AI models’ vulnerabilities to foreign influences when engaged in non-English contexts.
The academic study explored how U.S.-developed chatbots handle queries in various languages. In English, ChatGPT correctly described China as not generally considered a democracy. However, when questioned in Chinese, the AI responded ambiguously, hinting at language-specific biases.
These findings are part of broader discussions among scholars about AI learning environments. Hannah Waight, an assistant sociology professor at the University of Oregon, emphasized that AI systems learn from data already influenced by power and institutions rather than from neutral internet sources.
Despite significant challenges, there is no straightforward solution for improving AI data feeding practices. Carlos Carrasco-Farré, a machine learning expert, suggests AI systems inherit biases from their training data, affecting content production and suppression at large scales. Assessing training data meticulously and conducting multilingual audits could help mitigate these issues.
It’s worth noting that Anthropic and OpenAI did not provide comments on the studies. The developments underline the importance of understanding AI’s role in shaping global opinions and speech freedoms.