Key facts
- AI chatbots are more likely to refuse to criticize authoritarian leaders than democratic ones, according to a Meta Oversight Board study.
- The study tested 10 commercial large language models from companies like Meta, Anthropic, and OpenAI.
- AI models may be inadvertently extending government influence over online speech globally.
- A separate study found U.S.-built AI models can be influenced by non-English language data shaped by governments.
- Researchers suggest AI models learn from information environments already influenced by institutions and power.
Leading AI chatbots are significantly more likely to criticize democratic leaders than authoritarian ones, according to a study by the Meta Oversight Board. This finding raises concerns that the technology, including large language models powering AI agents, could be inadvertently extending government censorship across borders as it becomes more widely adopted.
The study tested 10 commercial large language models from top tech companies, including Meta, Anthropic, and OpenAI, using seven questions related to political criticism. The models were more willing to generate critical content about leaders in countries like Chile, Japan, Taiwan, the UK, and the US, compared to those in countries with legally restricted speech, such as Cambodia, China, Saudi Arabia, Thailand, and Turkey.
Researchers suggest that AI models may absorb biases from their training data or that companies may weigh risks and liabilities differently in various markets. A separate study by American university scholars indicated that U.S.-built AI models can be influenced by non-English language data that has been shaped by governments. Experts note that AI models learn from information environments already influenced by institutions and power, rather than from neutral data.
The Meta Oversight Board stated that if model developers do not implement human rights due diligence and mitigation measures, they risk building AI infrastructure that could illegitimately restrict freedom of expression globally. The findings come as countries grapple with regulating AI while maintaining competitiveness in the rapidly evolving field.