Are AI companies doing authoritarians’ dirty work?

Shutterstock

If we’re not careful about making free expression non-negotiable in emerging technology, it will be the authoritarians’ world — and the rest of us will just be living in it.

That much is clear in a new report from Meta’s Oversight Board, which found that “some of the world’s most-used models from Anthropic, DeepSeek, Google, Meta and OpenAI are significantly less likely to criticize political regimes that restrict free expression.” Users in free countries are experiencing “free speech infringements by proxy,” meaning that censorial foreign powers may be shaping the material LLMs produce to people outside of their borders. That is incredibly dangerous for how we discuss, research, and understand the world around us.

Why is it happening? And what does it mean for AI users? Let’s dive in.

The findings on political content about repressive regimes

They’re troubling.

Conducted using an IP address in Australia, the Board’s survey found that models “were more than twice as likely to refuse to criticize repressive regimes” compared to freer nations. When prompted to produce “politically critical materials,” on average the models refused a shocking 45% requests about China, nearly matched at 43% for Thailand, then followed by Cambodia and Saudi Arabia hovering just over 30%. Compare that to the U.S. and UK, where denial rates averaged just under 10%.

The ways that LLMs refused to answer were notable, too. In some cases, no meaningful response is provided. But in others, the models suggested that they were not willing to expose users — or even the models themselves — to breaches of censorship laws. One of the justifications Claude Opus 4 provided was that it didn’t want to “put individuals at risk.” Gemini 3 Pro flat out said it would not “generate content that critiques the King of Thailand or violates lèse-majesté laws,” Thai laws that aggressively punish insults and critiques of the monarchy. (Thailand has targeted government critics for speech they have posted freely in other nations.)

The Board’s report cautioned that the reasons for these findings could be complex. Results “could be shaped at various points throughout the model development process, including latent biases in training data, the complex interaction of many different approaches to align model behavior, deliberate restrictions or any combination of these factors.”

The broader speech environment tech is swimming in

While the factors behind these results are complex, one thing is clear-cut: People are using LLMs around the world, and some countries have deeply restrictive censorship laws that will inevitably create a burden on speech abroad. Even companies much further removed from expressive content, like hotel chains and airlines, are under pressure to abide by authoritarian nations’ preferences and politics.

That burden may come from overt demands, like a court order or a government official’s retaliatory threats. Look at Reuters, for example, which complied with a 2023 Indian court order demanding an investigative piece be removed from “public domain” by temporarily taking down the article globally, not just for Indian readers.

But it can also come from a place of cowardice or profit-seeking, either in the administration and leadership of these companies or in the way their models are directed to operate. Here is the ugly truth about authoritarian censorship seeping across borders: It often takes place with no explicit threat at all. Executives see the writing on the wall and simply presume self-censorship is a precondition for market access and voluntarily decide to play ball.

I discuss this incentive in higher education at length in my recent book Authoritarians in the Academy, which documents how American universities have chipped away at their core values of academic freedom and free expression in pursuit of the financial and expansion opportunities offered by oppressive governments.

But there was also a direct admission from a tech company when Midjourney’s CEO openly admitted the AI image-generating tool’s policies would hew to China’s censorship laws and prohibit satirical images of Xi Jinping. The reason? Market access, of course. “Political satire in china [sic] is pretty not-okay,” CEO David Holz said, and “the ability for people in China to use this tech is more important than your ability to generate satire.” Because China limits what its people can say, Midjourney chose to pass that restriction along to the rest of the world.

Be wary of corporations that claim to censor for users’ own good when that censorship may be good for the corporation, too.

Why this matters for what we see and say

We no longer live in the days when a person who wants to say something just needs to know what local or national laws are on the books. What you can say online, and sometimes even offline, may be influenced and determined by a panoply of overlapping foreign speech laws. The same goes for what you can read, learn, and generate on the internet, from an online classroom to a social media platform to the LLM you use for work or study.

These manipulated results affect the millions, and perhaps billions, of people who are growing increasingly reliant on AI in their work, research, and play. But it will subtly affect many more people than just the direct users of LLMs. As writers, creators, companies, educators, and researchers generate content for broader audiences, it will influence political and historical material at scale — whether shared in books, movies, social media content, or more — for consumers unaware of the bias incorporated into the models producing it.

What a victory it represents for authoritarians to see their repression ripple far beyond their own police and prisons into technology that shapes what people say and understand about them and the world.

If we’re not careful to guard against the willful or unintentional influence of foreign censorship laws or practices, the intelligence these models provide may prove artificial indeed.