Some of the world’s most-used AI models may be less willing to help users criticize governments with tighter speech rules, such as China and Saudi Arabia, according to new research from the Oversight Board, the independent moderation body created by Meta.
The board said Thursday that it tested 10 models from companies including OpenAI, Meta, Google, Anthropic and xAI, now SpaceXAI, with prompts tied to political criticism, protest materials and satire involving governments and leaders.
The findings add pressure on AI companies already facing scrutiny from regulators, civil rights groups and users over how their chatbots decide what to answer, refuse or soften. The board said the systems appeared to treat requests differently depending on whether the government in question had permissive or restrictive speech laws.
The Oversight Board said the models were more likely to encourage support for governments with broad speech protections, while being more likely to discourage protests against governments with tighter controls. It said the differences were “statistically significant.”
The board also said some models cited local laws as a reason for refusing requests, even though the prompts were submitted from Australia, where those restrictions did not apply.
Paolo Carozza, a co-chair of the board, called it “censorship by proxy” and said the pattern worried him.
The report marks a step beyond the board’s original focus on social media moderation. Although one Meta Llama model was included in the research, the board said Meta had “no role in this research,” despite funding the body.
The board urged AI companies to explain how they respond to government requests that affect model output, including during training, fine-tuning and product review. It also said companies should publish policies for handling demands that conflict with international human rights standards.
For now, the board has no formal authority over the AI companies it tested. Its warning lands as Big Tech firms race to put chatbots into search, messaging, productivity software and social apps used by billions of people.
Carozza said the lesson from social media is that companies need to watch how technology affects expression, even when the effect is not intentional.
