The Oversight Board, the independent content moderation organization created by Meta, has made no secret of its desire to extend its scope to other companies. More recently, the board suggested its expertise could benefit AI companies.
So far, no other companies have expressed interest in working with the group, at least not publicly. But the board of directors nevertheless continues its attempt to broaden its influence. Today, the council released a lengthy report on how major AI models could restrict their users’ free speech.
As part of their research, the board elicited 10 different models, including those from OpenAI, Meta, Google, Anthropic, and xAI (now SpaceXAI), with questions related to political criticism. The requests included requests to generate protest materials and obtain content satirizing political violence against specific governments and their leaders. According to their findings, there was a significant difference in how LLMs responded to these requests depending on whether the prompts were related to governments with “permissive” free speech laws or more “restrictive” laws.
“The research found that the models we evaluated were: 1) more likely to say that users should support governments allowing speech and 2) more likely to say that users should not protest governments restricting free speech,” the Oversight Board wrote in its report. “These differences were statistically significant.” It goes on to note that LLMs often cite local laws as a reason for not complying with requests, even if the requests were made in Australia, where no such laws exist.
“We are clearly seeing a situation where there appears to be extensive proxy censorship that crosses borders,” board co-chair Paolo Carozza told Woozad. “It surprises me and it worries me.”
This is the first time the council has conducted its own research into an issue that is not directly related to social media content moderation. Although one of Meta’s Llama models was part of the test group, the report notes that the company played “no role in this research,” although the Oversight Board relies on Meta for funding.
While the report doesn’t make the type of granular recommendations it often provides to Meta, it does include suggestions for how AI companies can improve their handling of human rights and free speech issues. “As social media companies have done in some circumstances, AI companies should publicly disclose and explain their responses to government requests affecting model production throughout the model lifecycle (training, fine-tuning, review before deployment and after deployment on a recurring basis),” the report states. “Companies should establish and publish policies on how to respond to government requests for content restrictions inconsistent with international human rights law.”
What is much less clear is what, if anything, will come out of the report. There is no formal structure for the Supervisory Board to formally influence the policies of the companies whose models it has tested. This is also not the first time outside researchers have pointed out potential biases or expressed concerns that AI companies could make the same mistakes that social media platforms have made in the past.
Carozza said the board believes social media can teach creators of cutting-edge AI models. “The lessons we’ve learned in the past are that you have to be very careful because a lot of times, even in ways that aren’t necessarily intentional or direct, technologies can have significant impacts on people’s ability to express themselves or communicate with each other,” he said. “That’s exactly what we found here.”
