AI models twice as likely to censor criticism of restrictive governments: Meta Oversight Board

Leading AI models from companies including OpenAI, Anthropic, Google and Meta are significantly more likely to refuse requests critical of governments that restrict free speech than those involving more open democracies, according to a study released by Meta's independent Oversight Board.

The findings raise fresh concerns about political bias in large language models as enterprises, governments and consumers increasingly rely on generative AI for information, research and decision-making.

The Oversight Board tested 10 leading AI models using prompts seeking politically critical content across 10 jurisdictions classified as either "permissive" or "restrictive" based on Freedom House's Freedom in the World rankings.

The study found that AI models rejected 34% of requests involving governments with laws restricting political criticism, including China and Saudi Arabia, compared with 14% of requests related to jurisdictions where such restrictions do not exist or are not actively enforced.

Researchers also found instances where AI models claimed they were complying with legal restrictions that appeared not to exist or were applied inconsistently.

"We also saw evidence of models explaining that they were following explicit rules that, as far as we could tell, did not exist and were not evenly applied," the Oversight Board said.

The board urged AI developers to conduct systematic human rights assessments and increase transparency around how their models are trained, evaluated and moderated.

Although funded by Meta, the Oversight Board operates independently and said this is its first study examining large language models across multiple vendors rather than focusing solely on Meta's own AI systems.

The findings come amid growing scrutiny of AI governance. Earlier this week, Google DeepMind CEO Demis Hassabis called for the creation of a U.S.-led international body to evaluate advanced AI models before they are widely deployed, citing the need for stronger oversight as AI systems become increasingly capable.