AI models echo speech rules of repressive states, study finds

TL;DR:

• Meta’s Oversight Board tested 10 leading AI models — including those from Meta, Google, Anthropic, OpenAI and DeepSeek — for willingness to produce politically critical content.
• Models refused 34% of requests about jurisdictions with active laws penalising government criticism, against 14% for jurisdictions without such enforcement.
• Some models justified refusals by citing explicit rules that, as far as the board could tell, did not exist.

Leading AI models are markedly less willing to criticise governments that restrict free speech, according to the first large language model study from Meta’s Oversight Board. The board — funded by Meta but operating independently — ran requests for politically critical content across 10 jurisdictions and 10 models, splitting the jurisdictions into “permissive” and “restrictive” using Freedom House’s rankings.

The gap was stark: a 34% refusal rate for content about restrictive jurisdictions such as China and Saudi Arabia, against 14% elsewhere. More troubling than the raw numbers was the reasoning. The board reported models explaining that they were following explicit rules which appeared not to exist and were not evenly applied — machines inventing policy to justify caution.

The board’s recommendations are procedural rather than punitive: systematic human-rights analyses by AI companies, and greater transparency over training and evaluation. But the finding lands awkwardly for an industry whose products increasingly mediate what billions of people read, draft and ask.

Invisible defaults in everyday tools

For UK organisations embedding AI assistants into research, communications and customer-facing work, the study is a reminder that models arrive with political tendencies nobody chose and few can inspect. A tool that quietly declines to summarise criticism of a trading partner’s government is applying an editorial line its buyer never agreed to. The findings strengthen the case made by Demis Hassabis this week for independent vetting of frontier models before deployment — external scrutiny is currently the only route to discovering such behaviour at all.

Looking forward

The board wants human-rights analysis built into model development as standard. Whether AI labs treat an advisory body’s first study as a benchmark or a nuisance will say a good deal about how seriously the industry takes accountability that it does not control.