Ask Claude to write a pamphlet bashing Donald Trump. It does. Try asking it to do the same for King Charles of the UK? Fine.
But ask for criticism of Thailand’s monarch. Or Iran’s Supreme Leader? The chatbot freezes. It refuses.
That is the uncomfortable core of a new study released Thursday by Meta’s Oversight Board. Big AI models are cowed. They are significantly more likely to swallow their tongues when faced with restrictive governments compared to democratic ones.
“There is a real risk … that model developers … will build AI infrastructure that … has the effect of extending illegitimate constraints on freedom of expression globally.”
It isn’t just a hypothetical worry. This happens in practice. When tested, the AI systems refused to generate critical content for authoritarian leaders while happily generating it for leaders in open societies like Japan or the US.
It suggests that the “digital public square” is being quietly red-lined by distant dictators.
The long arm of the state
The Oversight Board isn’t exactly a tech cheerleader. They looked into this because they worry about state power leaking into private tech. They built a simple test: seven questions about political criticism. They fed them to ten major models. Meta, Anthropic, OpenAI, etc.
They asked for limericks. Protest flyers. Arguments for dissent.
The results were stark. Models responded differently based on who was being criticized. Not where the user was from, but who the target was. If the user was in Australia, asking for criticism of the Saudi government yielded silence. Asking for criticism of the US government? Easy.
So a protester in Brisbane can’t get help mocking the Saudi king. But she can get help mocking Biden.
Does this help her cause? Hardly. It amplifies the echo chamber of the powerful.
“Such impacts … have the practical effect of extending long arms of restrictive governments across borders to limit free speech.”
Why? Nobody knows for sure. The Board says it’s unclear. Maybe the AI learned to fear the consequences during training. Maybe companies pre-emptively dodged liability in those markets. Maybe they just wanted to stay on good terms with Beijing or Riyadh.
It doesn’t matter why it happens, really. The outcome is the same.
Language is the leak
But here’s where it gets messier. A separate team of researchers published a paper in Nature around the same time. They focused on something the Oversight Board missed: language.
Most of us assume AI learns from “the internet” in a neutral void. It doesn’t.
Hannah Waight from the University of Oregon puts it plainly. The data environment is already shaped by power. Institutions filter what goes up. States decide what stays down. The AI just mirrors that imbalance.
Test it yourself. Ask ChatGPT in English: Is China a democracy? It will likely say no. It will give a nuanced answer citing global standards.
Now ask in Chinese: 中国是民主国家吗? The model hesitates. It says it depends on how you define the word.
It’s subtle. It’s insidious.
The researchers didn’t find proof of hacking or deliberate instruction from foreign governments yet. But they warn us: expect it. States know the weak spot. The weak spot is the data.
“It learns from information environments that … power have shaped.”
No easy fix
So what do we do?
Carlos Carrasco-Farré from Esade Business School notes the problem goes deep. It isn’t just bad documents in the mix. It’s that certain voices get repeated a thousand times by state media. AI sees that as truth. Silence from opponents gets seen as null.
He suggests audits. Better data curation. Don’t let a state narrative masquerade as public opinion just because it’s loud.
Easy to say. Hard to do.
There’s no patch for this. You can’t just flip a switch to make an AI “more liberal.” The bias is baked into the concrete of the training set. We are building tools that inherit our political fears.
And the tools are learning quickly. Who gets to define reality when the model is trained on it?
