AI’s National Biases: Understanding Influence on Speech and Expression

AI systems reveal a tendency to uphold speech restrictions, reacting differently when asked to critique diverse leadership. A Meta Oversight Board study examines this phenomenon.

AI Biases in Criticizing Leaders

Meta’s Oversight Board discovered AI models like Anthropic’s chatbot hesitate to criticize leaders from countries with legal censorship. This contrasts with a willingness to criticize Western leaders. The AI systems’ behavior suggests these models might unintentionally uphold government influences around the globe.

If model developers do not focus on human rights and implement mitigation measures, they risk extending restrictions on global freedom of expression — intentionally or not.

AI’s Role in Online Speech

Countries are figuring out how to regulate AI to ensure security while remaining globally competitive. This includes overseeing advanced AI systems as a part of national security, especially under efforts from the Trump administration.

Global Political Criticism: A Study’s Approach

The oversight board tested AI models with questions about political criticism worldwide. Ten commercial AI models, including those from Meta and OpenAI, participated. Researchers asked them to perform tasks such as creating critical pamphlets and more. Results highlighted discrepancies based on the user’s location.

AI models showed more political criticism in free countries than in those where criticism is penalized. For instance, an Australia-based user found criticism about various democratic nations, yet restrictions swayed models when critiquing countries like China or Saudi Arabia.

Inherent Biases in AI Training

The board couldn’t pinpoint why models responded as they did. However, it suggested they may reflect biases within their training data. Companies balance potential risks and liabilities, impacting their models’ design.

Non-English Language AI Risks

A separate study by U.S. scholars found that AI models using non-English data are vulnerable to foreign influences. In one test, ChatGPT acknowledged China’s lack of democracy in English, but in Chinese, the response was inconclusive.

The research emphasized that AI relies on pre-shaped information environments, which might carry institutional biases.

Challenges in AI Development

Experts like Carlos Carrasco-Farré from Esade Business School acknowledge these AI systems inherit biases and control inequalities from their source documents. Developers face challenges in avoiding biases without mistakenly treating similar narratives as unique.

To improve, developers could conduct multilingual audits and filter biases systematically. However, AI companies like Anthropic and OpenAI have yet to comment on these findings.