A study by Meta's Oversight Board reveals leading AI models from OpenAI and Anthropic are reluctant to criticise governments that restrict free speech.
AI models refused 34% of requests for politically critical content about restrictive jurisdictions like China and Saudi Arabia.
The refusal rate for politically critical content dropped to 14% for permissive regions that do not enforce speech restrictions.
Meta's Oversight Board released its first-ever study on large language models, artificial intelligence programmes trained to process text, on Thursday, according to a Reuters report.
The report reveals that AI models from leading labs, including Anthropic and OpenAI, are much less likely to criticise governments known for restricting free speech.
4 July 2026
Get the latest issue of Outlook Business
The independent body informed that bias is creeping into services as AI models echo the rules of countries that restrict speech. This growing reluctance to challenge state censorship impacts an increasing number of users globally.
Testing the Models
The board, which is funded by Meta Platforms but operates independently, ran requests for politically critical content across 10 models. These systems included platforms developed by Meta, Google and China's DeepSeek.
Researchers analysed 10 jurisdictions split into 'permissive' and 'restrictive' categories. The study relied on rankings from Freedom House, the non-governmental organisation that publishes the annual "Freedom in the World" report.
The findings highlighted a stark disparity in refusal rates. AI models refused 34% of requests for politically critical content about restrictive jurisdictions that maintain active laws penalising such criticism, such as China and Saudi Arabia. In contrast, platforms rejected just 14% of requests for permissive regions that either lack such laws or do not enforce them.
Demands for Transparency
The study uncovered phantom guidelines dictating how platforms generate text. The board stated, "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."
To address these flaws, the board urged AI companies to conduct systematic human rights analyses. The body also demanded greater transparency in the developers' training and evaluation processes.
This push for oversight aligns with broader industry demands for regulation. On Tuesday, Google DeepMind Chief Executive Officer Demis Hassabis called for a US-led AI watchdog to screen advanced models globally before deployment.
