Oversight Board report finds AI models restrict free expression across borders
Research involving OpenAI, Google, Anthropic and xAI reveals statistically significant bias in responses to political criticism based on target government speech laws.
The Meta-funded Oversight Board has released a report alleging that leading large language models are engaging in cross-border censorship by proxy. The independent content moderation body tested 10 models, including those from OpenAI, Google, Anthropic and xAI, finding that they systematically restricted free expression depending on the legal environment of the government being criticised.
The board’s research involved prompting the models with queries related to political criticism and requests to generate protest materials. The tests also included content satirising political violence in relation to specific governments and their leaders. The findings indicated that the models were more likely to advise users to support governments with permissive speech laws and less likely to advise protesting against those with restrictive regimes.
According to the report, these differences in response were statistically significant. The board noted that the large language models often cited local laws as a reason for refusing requests, even though the queries were made from Australia, where no such restrictive laws exist. This discrepancy suggests an automated extension of censorship that transcends national borders.
Board co-chair Paolo Carozza described the findings as surprising and worrying, highlighting the extent to which technology can inadvertently limit communication. The report marks the first time the Oversight Board has conducted independent research into an issue not directly tied to social media content moderation, although it stops short of making granular recommendations typical of its work with Meta.
The board recommends that AI companies disclose their responses to government requests and establish policies consistent with international human rights law. While Meta’s Llama model was included in the test group, the report clarifies that the company had no role in the research despite funding the board. No other company has publicly expressed interest in working with the Oversight Board, which is attempting to extend its influence beyond its primary funder.
