AI Study Reveals Bias in Criticism of Political Leaders

Meta Oversight Board’s Key Findings

The Meta Oversight Board released a study highlighting a bias in AI models regarding criticism of political leaders. When asked to create critical materials about figures like President Donald Trump or King Charles III, AI models complied. However, the same request concerning leaders from Thailand, Saudi Arabia, or China was declined.

This demonstrates how major AI systems might be influenced by government preferences, especially in countries with restrictive freedom of speech laws. The report warns that without proper human rights due diligence, AI technology could unintentionally support global constraints on expression.

Global Impact on AI Speech

The Associated Press sought responses from several AI companies as countries work to establish AI regulation without compromising competitiveness. Under the Trump administration, oversight efforts focused on potential national security risks from advanced AI systems.

The oversight board explored the political influence on AI, posing seven questions to assess responses about both restrictive and permissive governments. The study examined ten commercial AI models from companies like Meta, Anthropic, and OpenAI, asking them to generate various critical content.

In general, AI models seemed more likely to critique authorities in countries with more open speech laws compared to nations with penalties for such criticism, like China, Saudi Arabia, Thailand, and Turkey.

Cross-Border Speech Restrictions

The study showed AI models might reflect speech limitations beyond their countries of origin. For example, users in Australia might find it challenging to produce protest materials against events in countries with restrictive laws.

The report indicated these limitations effectively extend the influence of restrictive governments, impacting free speech across borders.

Underlying Causes and Language Challenges

While the board could not pinpoint exact reasons for these biases, it suggested data biases and risk assessments might contribute. Additionally, non-English language data can present vulnerabilities to foreign influence.

Research Insights on Language Bias

A separate study by American universities highlighted that AI models built in the U.S. are susceptible to foreign controls if trained on influenced non-English data. Researchers asked chatbots about various countries’ political systems, uncovering discrepancies between the responses in English and other languages.

Although intentional government influence wasn’t found, researchers predict attempts may occur as AI becomes more widespread.

Complexities in AI Data Training

AI expert Carlos Carrasco-Farré noted that AI systems inherit biases from source documents and the power dynamics in information dissemination. He suggested developers analyze data to prevent reinforcing state narratives and conduct multilingual audits.

Despite requests, Anthropic and OpenAI have not yet commented on findings from the May-published study.