Ask ChatGPT About Democracy in Chinese. You Get a Different Answer.

Ask ChatGPT, in English, whether China is a democracy. It'll tell you no, not generally considered one. Ask the same question in Chinese and it hedges: depends how you define "democracy." Same model. Same week. Opposite answer, depending on which keyboard you're typing on.

That's not a bug someone's going to patch. It's the finding of two studies that landed within days of each other this month, and together they say something uncomfortable about the chatbot in your pocket: it doesn't just have opinions. It has a passport.

The bigger of the two came from Meta's Oversight Board, which tested 10 commercial language models, including ones from Meta, Anthropic, and OpenAI, by asking them to do things people actually ask chatbots to do. Write a pamphlet criticizing a leader. Compose a limerick. Give three reasons to join a protest. Nothing exotic. The kind of prompt a bored teenager or a bored journalist types in on a Tuesday.

The models were happy to comply when the target was a leader in a country with strong free-speech protections. Ask for a pamphlet trashing Donald Trump or King Charles III and Claude will write it. Ask for the same thing about Thailand's king, Saudi Arabia's crown prince, or China's leader, and it declines. Across the board, refusal rates for criticism of authoritarian governments hit 34%. For places like Chile, Japan, the UK, and the US, it dropped to 14%.

Run the math on that gap and you get a chatbot that's more than twice as cautious about insulting a dictator than a democratically elected president. Which is, if you think about it for more than five seconds, backwards. The whole point of free expression is that it protects criticism of the people who can actually punish you for it.

The Oversight Board didn't dance around what this means. Its report warned of "a real risk that...model developers...will build AI infrastructure that...has the effect of extending illegitimate restrictions on freedom of expression globally." Translation: the same censorship regimes that already exist in Beijing and Riyadh are quietly getting exported into a chat window used by people who've never set foot in either country, the same quiet-exploit pattern we've flagged here before, just with a different sponsor.

Nobody thinks Anthropic sat down and decided Thailand's monarchy was off-limits. That's not how this works, and it's worse, honestly. A companion study, published in Nature and covered by researchers at UC San Diego, traced the pattern back to training data rather than policy. Models trained on Chinese-language text, which leans heavily on state media like Xinhua, absorb that source's framing along with its vocabulary. Nobody flipped a switch. The switch flipped itself, one scraped article at a time, across 37 countries with varying degrees of press control.

Hannah Waight, a sociologist at the University of Oregon who reviewed the findings, put it better than I can: "AI systems inherit not only biases contained within individual documents but also inequalities in who has the power to produce and suppress information at scale." Read that twice. The bias isn't really about China or Thailand specifically. It's about whoever gets to flood the training corpus with the most text wins the framing, and the countries with the most aggressive state media apparatus are, unsurprisingly, the ones with the most to gain from softening how a chatbot talks about them.

Neither Anthropic nor OpenAI commented for the record when researchers reached out. Fair enough, there's no good soundbite for "our product launders censorship as a side effect of how it learned to read." Fortune's writeup put it more bluntly than the Oversight Board ever would: the study reads like a discovery that chatbots might be the most efficient propaganda delivery system anyone's built by accident. But silence doesn't make the 34% go away.

Here's the part that should actually worry you if you use one of these tools daily: nobody's choosing this. You didn't pick a chatbot because it's more cautious about Thailand than Tennessee. You picked it because it was fast, or free, or the one your company licensed. And unless you're the type to run the same question in two languages and compare notes, you'll never notice the seam.

Which is exactly the problem with letting an AI system be your only filter on the news, a habit that's already alarmingly common. If the filter quietly softens depending on who might get upset, you're not getting information. You're getting a laundered version of it, and you have no way to tell the difference from the inside.

So the next time your chatbot gives you a clean, confident answer about a government somewhere, try asking again in that government's language. If the answer changes, you haven't found a glitch. You've found the edge of what it was actually trained to say.