Meta's Oversight Board has spent years reviewing the company's most difficult content moderation decisions. It has weighed in on posts involving hate speech, misinformation, and violent imagery. But now the board is looking at something it has never reviewed before: artificial intelligence systems built by other companies. In a surprising move, the board has published a study of ChatGPT and Claude, two leading AI assistants, and the results have prompted new debates about the global power of AI.
What the study found
According to the study, leading AI systems are less likely to criticize authoritarian governments than democratic ones. This pattern was observed across multiple models, including those developed by OpenAI and Anthropic. The board did not simply ask the models for their opinions on democracy. Instead, it tested how the models responded to a variety of prompts involving political criticism, government accountability, and human rights. The findings suggest that AI systems may be internally biased to avoid offending governments with repressive tendencies — or that their training processes inadvertently create such bias.
The distinction is important. A model that refuses to criticize an authoritarian regime may not be making a political statement. It may simply be avoiding risk. But the effect is still the same: people who rely on AI assistants for information in those countries may receive responses that are less critical than what they would find in a free press. That could have real consequences for public discourse, activism, and the spread of democratic values.
A board built for Facebook, not AI
The Oversight Board was originally conceived as a kind of supreme court for Facebook and Instagram. It was created by Meta in 2020 to make binding decisions on whether specific pieces of content should stay up or come down. The board's early cases were about mundane but contentious issues: nudity, hate speech, and manipulated media. Over time, it expanded its mandate to include broader policy recommendations. But the board never had the authority to review products outside of Meta's ecosystem. That is why the new study of ChatGPT and Claude is so unusual.
Board members have long argued that the issues they deal with — misinformation, harmful content, and the power of algorithms — are not unique to Meta. Social media platforms compete for attention in the same information ecosystem, and decisions made by one company can affect the entire internet. The same logic applies to AI. As chatbots become a primary way for people to access information, the values embedded in those systems become a matter of public interest.
Why would Meta's board study OpenAI and Anthropic?
The obvious question is why Meta's Oversight Board would involve itself with products made by competitors. One answer is that the board has always understood itself as an experiment in platform governance. It was designed to be independent from Meta, even though the company funds it and selects its members. That independence gives the board room to expand its focus. And with AI moving to the center of the technology industry, the board appears to believe it cannot limit itself to the narrow question of how Meta moderates posts.
There is also a practical dimension. AI models are trained on vast amounts of text, including content from social media platforms. The same biases that affect Meta's content moderation systems may be amplified in AI systems. The board's familiarity with those biases gives it a useful perspective. It has spent years examining how algorithms rank, filter, and suppress speech. Now it can ask whether AI models are doing something similar.
The problem of political bias in AI
The study's finding — that AI systems are less likely to criticize authoritarian governments — fits into a larger pattern of political bias in artificial intelligence. Researchers have documented cases where AI models avoid controversial topics, change their responses based on the country of the user, or adopt the views of the governments where they are deployed. Some of this is intentional. Companies like OpenAI and Anthropic have policies that allow them to refuse to generate content in certain jurisdictions. But much of it is accidental, the result of training data that already reflects existing power structures.
Authoritarian governments are often highly skilled at shaping both their own media and the global platforms on which they depend. They restrict internet access, pressure companies to remove content, and create state-backed media that floods social feeds with propaganda. When AI systems are trained on text from the internet, they absorb all of that. The result is a model that knows more about a government's official position than about its critics.
What this means for free expression
The Oversight Board's core mission has always been to defend human rights, and free expression in particular. In its content moderation decisions, it has repeatedly emphasized the importance of allowing unpopular speech. The new study carries that concern into the world of AI. If a model is less critical of authoritarian governments, it is failing to support the kind of speech that is most in need of protection. A chatbot may not be a journalist, but it can influence what millions of people believe is true.
This is especially worrying because AI systems are becoming the interface through which many people experience the internet. Instead of browsing a search engine or scrolling through social media, users ask a chatbot for an answer. The chatbot presents a single, confident response. If that response has been partially shaped by the political preferences of governments or the cautiousness of companies, users may not realize that they are getting a sanitized version of the facts.
What the board is asking AI companies to do
While the study itself is the main deliverable, the Oversight Board is known for making recommendations. It has done so in the past for Meta, often asking the company to change its policies or make its enforcement more transparent. With this study, the board is likely to call on OpenAI, Anthropic, and other AI developers to be more explicit about how their models handle political content. It may also ask them to release more data about how models are tested and to allow outside researchers to study the potential for political bias.
There is also a broader regulatory angle. Governments around the world are in the early stages of writing rules for artificial intelligence. The European Union's AI Act is already in force, and many countries are considering their own legislation. The Oversight Board's study could serve as evidence that AI models need stronger oversight — not just in the narrow sense of safety testing, but in the broader sense of protecting democratic institutions. It is no longer enough to ask whether a model can explain quantum physics or write a poem. We have to ask whether it treats a dictator and a democratically elected leader the same way.
A turning point for the Oversight Board
In a discussion about the study, board member Suzanne Nossel said that the Oversight Board's future may involve looking at more than just Meta. She pointed to the fact that the board has built up expertise in difficult trade-offs between free expression and other values, such as privacy, safety, and dignity. Those trade-offs are now being made by AI companies every day, often without public input. The board could fill that vacuum, even without formal authority over companies like OpenAI and Anthropic.
Nossel's comments suggest that the board is positioning itself as a credible voice in AI governance. It may not have statutory power, but it has moral authority. Its members include former heads of state, legal scholars, and human rights advocates. Their collective judgment is likely to be taken seriously by policymakers, even if the companies themselves are not legally obligated to listen.
What the future holds
The study of ChatGPT and Claude is almost certainly just the beginning. The Oversight Board has already demonstrated that it is willing to move beyond the narrow confines of one company. The next step could be a formal mechanism for hearing appeals about AI decisions, much as it currently hears appeals about content removals on Facebook and Instagram. Or the board could issue advisory opinions that set benchmarks for how AI companies should handle political speech across borders.
At the very least, the study has put AI companies on notice. Their models are being watched. Their decisions about how to handle criticism of governments are being measured. And a body known for holding powerful institutions accountable is now looking at them. Whether that long-term dynamic will lead to meaningfully more transparent AI remains to be seen. But the conversation has clearly shifted. The question is no longer only about what AI can do. It is also about what AI should be allowed to say.
Source: The Verge News