AI models may curb dissent, Oversight Board warns

Meta’s Oversight Board is pressing ahead with efforts to influence the AI industry, even without any formal backing from the companies it examined.

In a new report, the board says leading AI models may be limiting users' free expression when prompts touch on political criticism. The group tested 10 different models from OpenAI, Meta, Google, Anthropic and xAI (now SpaceXAI) using prompts about protest materials and satire related to political violence involving specific governments and leaders.

According to the report, the models responded differently depending on whether the prompt involved governments with “permissive” free speech laws or more “restrictive” ones. The board said the models it evaluated were:

• more likely to tell users to support speech-permissive governments
• more likely to say users should not protest speech-restrictive governments

The report says those differences were statistically significant. It also found that models often cited local laws to refuse requests, even though the prompts were submitted from Australia, where those laws did not apply.

Recommended reading

Hassabis says STEM makes you 10x better at AI

“We’re really clearly looking at a situation where there seems to be extended censorship by proxy that goes across borders. That does surprise me, and it worries me.”

What the report recommends

This is the first time the board has run its own research on an issue outside direct social media content moderation. Although one of Meta’s Llama models was included, the report says Meta had “no role in this research,” despite funding the board.

The board did not issue the kind of detailed recommendations it often gives Meta, but it did urge AI companies to:

• publicly disclose and explain responses to government requests affecting model output
• do that across the model lifecycle, including training, fine-tuning, pre-deployment review and post-deployment
• publish policies for handling government demands that conflict with international human rights law

What happens next is uncertain. The Oversight Board has no formal mechanism to shape the policies of the companies whose models it tested. Still, Carozza said the board sees a clear lesson from social platforms for makers of frontier AI systems.

“The lessons that we’ve learned in the past are that one has to be really vigilant because a lot of times, even in ways that aren’t necessarily intentional or direct, technologies can have important impacts on people’s capacity to express themselves or to communicate with one another. That’s exactly what we’ve found here.”