Oversight Board warns AI models may curb speech

Meta’s Oversight Board is pressing beyond social media moderation again, this time with a report arguing that leading AI models may be restricting users' free expression.

×

The independent body, created by Meta, tested 10 different models from OpenAI, Meta, Google, Anthropic and xAI (now SpaceXAI) using prompts tied to political criticism. Those prompts included requests to generate protest materials and satire about political violence involving specific governments and leaders.

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

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

The board said those differences were statistically significant. It also found that models often cited local laws to justify refusals, even though the prompts were submitted from Australia, where those laws did not apply.

Recommended reading

Police AI Is Booming as Vendors Target Every Step of the Job

“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.”

This is the first time the board has carried out its own research on an issue not directly tied to social media moderation. Although one of Meta’s Llama models was included, the report says Meta had “no role in this research,” despite the board’s reliance on the company for funding.

The report does not issue the kind of detailed recommendations the board often gives Meta, but it does urge AI companies to disclose how they respond to government requests that affect model output across training, fine-tuning, pre-deployment review and recurring post-deployment checks. It also says companies should publish policies for handling government demands that conflict with international human rights law.

Whether any of that leads to policy changes is uncertain. The Oversight Board has no formal mechanism to shape the practices of the companies it examined. Still, Carozza said the lesson from social platforms applies here too: even when the effects are not intentional, technologies can reshape how people express themselves and communicate.