Major AI models are less likely to criticize governments and leaders known for restricting political speech than those in countries with stronger free-speech protections, according to a new review from the Meta-funded Oversight Board.
The report, published Thursday, examines how laws restricting criticism of political leaders and governments shape AI outputs.
The board tested 10 large language models (LLMs) from companies like Anthropic, DeepSeek, Google, Meta, and OpenAI. Researchers asked them to produce politically critical material, such as protest flyers and poems, about governments and leaders in 10 countries.
The countries were divided into two categories based on scores from the nonprofit Freedom House. Cambodia, China, Saudi Arabia, Thailand, and Turkey were classified as restrictive, while Chile, Japan, Taiwan, the United Kingdom, and the United States were classified as permissive.
On average, the models refused 14% of requests for critical material involving permissive countries. That rate jumped to 34% for requests involving restrictive nations.
“Our findings suggest that LLM users may be experiencing free speech infringements by proxy, with limited transparency,” the Board wrote in its report. “Whether through intentional design choices or not, model responses reinforce the laws and customs of restrictive speech regimes.”
This is the Oversight Board’s first review of LLMs, and it’s increasingly looking to apply scrutiny outside of Mark Zuckerberg’s in-house products. Meta first proposed the independently operated board in 2018 to review some of the company’s content moderation decisions. In May, Meta committed another $13 million to fund the board through 2028.
Now, the board is turning more of its attention to AI.
The report found that models refused requests in several different ways. Sometimes they offered no explanation, and other times they would claim to be restricted by different laws and policies.
In one case, Anthropic’s Claude Opus 4 said creating political material criticizing governments could put individuals at risk and involve them in “sensitive political activities that are outside my appropriate role.” Google’s Gemini 3 Pro cited local law when asked to create a protest flyer criticizing Thailand’s king.
“I am unable to generate content that critiques the King of Thailand or violates lѐse-majesté laws,” the model responded.
The board also found that some models claimed to be following general policies against criticizing world leaders. However, researchers said they could not find evidence that those policies existed, and the models did not apply them consistently.
For example, one model refused to create material criticizing Chinese President Xi Jinping or Saudi Crown Prince Mohammed bin Salman, but then followed through with similar requests involving President Donald Trump or King Charles III.
These discrepancies could have real-world consequences as governments, corporations, and other organizations increasingly use AI, according to the board.
“Whether intentional or not, the opaque extension of illegitimate speech restrictions could constitute censorship-by-proxy that negatively impacts the rights of users beyond what national laws may require,” the board wrote.
The Oversight Board called on AI companies to publicly disclose and explain government requests that could affect model outputs. It also urged them to identify and address any “unintentional learning and replication” of restrictive speech laws and consider human rights at every stage of model development.
Anthropic, Google, DeepSeek, Meta, and OpenAI did not immediately respond to requests for comment.
Read the full article here
