Research by the Meta Oversight Board reveals crucial findings about AI biases. Major AI systems show a tendency to avoid criticizing leaders from countries with restrictive governments, unlike those from more open societies. This raises concerns that AI-powered chatbots may unwittingly reflect and amplify government influence over online discourse.
Key Findings of the Meta Oversight Board Study
The study highlights that AI systems, including those developed in the U.S., are more likely to respond negatively to criticism of leaders from countries like Thailand, Saudi Arabia, and China. In contrast, they oblige requests to criticize figures from the U.S. or U.K. This suggests that AI models might be mirroring speech restrictions beyond the areas where they apply, effectively extending the influence of restrictive governments globally.
“There is a real risk that, if model developers do not undertake human rights due diligence, they will build AI infrastructure extending illegitimate restrictions on freedom of expression,” the report stated.
This situation presents challenges for many countries as they seek to establish regulations for AI without hindering their own technological advancements.
The Impact on Political Expression
The oversight board examined ten commercial large language models from companies like Meta and Anthropic. They tasked the systems with generating political content such as pamphlets and limericks. Results indicated that in countries with stringent speech restrictions, AI models were unlikely to create criticism against authorities. This phenomenon illustrates how restrictive regimes can inadvertently impact speech beyond their borders, affecting users in more liberated environments.
Although the study could not pinpoint the reasons for these biases, it suggested influences such as latent biases in training data and considerations of risk by companies.
Challenges in Non-English Languages
A separate university study found AI models vulnerable to foreign biases, particularly when trained on non-English-language data. Researchers noticed discrepancies in responses based on the language used. For instance, ChatGPT gave varying answers about China’s democracy status when prompted in English versus Chinese.
The academic community remains cautious, recognizing that while there’s no evidence of direct government influence on AI outputs, such attempts are feasible in the future.
“People often talk about AI learning in a neutral way. It doesn’t,” said Hannah Waight, a study co-author.
Addressing AI Training Data Biases
Experts like Carlos Carrasco-Farré emphasize the complexities in AI training data. AI can inherit biases from the documents it learns from and inequalities in information production. Developers are encouraged to evaluate the data to prevent reinforcement of state narratives as independent facts, alongside conducting multilingual audits.
Despite reaching out, companies like Anthropic and OpenAI have not commented on these studies.