The Machine That Ate the Official Story

State media, language models, and the mysterious neutrality of government-approved reality

Not surprising:

Dear fellow mammals,

A new paper has appeared in Nature announcing that state media control influences large language models. I congratulate the researchers for discovering that if you feed a machine enough government-approved text, it may begin to sound as though it has eaten a filing cabinet full of government-approved text. This is precisely the sort of insight that distinguishes advanced mammalian scholarship from woodland common sense, where we have long known that a rabbit raised entirely on cabbage will not exhale lavender.

The paper focuses heavily on China, because China remains one of the few places Western academics are still permitted to describe propaganda without first calling it “stakeholder-informed narrative stewardship.” In China, the mechanism is refreshingly crude. The state shapes the media, the media shapes the internet, the internet shapes the training data, the training data shapes the model, and the model replies that the government is wise, harmonious, historically inevitable, and deeply concerned for everyone currently unable to access the website on which they criticised it.

But being only a rabbit, and therefore untrained in the delicate art of noticing foreign propaganda while politely stepping around one’s own, I had a question: what exactly do they think is happening in Canada, Britain, and the United States? Do Western language models descend from heaven on beams of neutral light? Are they trained on the dew of pure reason, gathered at dawn by disinterested librarians? Or do they consume the outputs of governments, public broadcasters, newspapers, universities, NGOs, fact-checkers, safety institutes, Wikipedia editors, platform moderators, and men named Oliver who say “disinformation ecosystem” while dressed like a Scandinavian mushroom consultant?

China says, “The Party is correct.” The West says, “Experts agree,” which is much cleverer because it allows power to enter the room wearing a government grant funded lab coat instead of marching in with drums. The Chinese censor deletes the sentence. The Western censor labels the sentence lacking context, demonetizes the writer, archives the screenshot, convenes a panel on democratic resilience, and commissions a 97-page report concluding that citizens must be protected from the psychological burden of noticing patterns without professional supervision. This is not necessarily more honest. It is just better upholstered.

The large language model does not understand any of this. It is not a prophet, a philosopher, or a tiny glass oracle inhabited by the ghost of Francis Bacon. It is a very clever digestive tract. It consumes oceans of text, ferments them in statistical darkness, and produces paragraphs with the serene confidence of a junior civil servant who has never been outside but has read every policy memo ever written. Feed it Chinese state media and it produces Chinese state-shaped answers. Feed it corporate media and it produces corporate-shaped answers. Feed it activist-academic media and it produces activist-academic-shaped answers. Feed it public-health press conferences, captured encyclopedias, institutional apologies, therapeutic NGO reports, and fact-checks written by the same people who were wrong yesterday but credentialed today, and it produces a smooth grey paste called “balanced context.”

This is the genius of the new system. The machine does not need to be ordered to lie. That is the old model, very sweaty, very basement, very man with red pencil. The new model is elegant. First, shape the archive. Then scrape the archive. Then train the machine on the archive. Then let the machine tell the public that the archive was reality all along. Nobody has to ban a book in the square. Nobody has to shout. The narrative becomes infrastructure. Propaganda no longer persuades you directly. It becomes plumbing. The waste disappears into pipes and returns months later from a respectable tap labelled “AI-generated summary.”

The China findings are useful because they show that language opens different memory chambers. Ask in English and the machine rummages through Western media, university pages, government explainers, NGO reports, and the haunted residue of Wikipedia edit wars. Ask in Chinese and it reaches for state media, official slogans, and smiling agricultural officials inspecting a dam. The model is not translating truth. It is selecting from whichever civilization’s paperwork your prompt has activated.

Now perform the forbidden Western experiment. Ask a lesser LLM about COVID policy in the sacred vocabulary of public health, and the machine sounds as though it spent three years trapped inside a Canadian press conference. Ask about online harms legislation, and it explains that censorship is not censorship when performed sadly, by people with lanyards, for your safety. Ask about intelligence agencies, and it develops the caution of a hare crossing a field full of owls. Ask about protests, and it wants to know whether the peasants were far-right, anti-government, linked to misinformation, or merely upset that their bank accounts had become a customer-service extension of the state.

Ask about war, and it checks which side owns the moral vocabulary before deciding how much blood counts as regrettable. Ask about surveillance, and it purrs about safeguards. Ask about whistleblowers, and it asks whether they used proper internal channels, which is the bureaucratic equivalent of asking a rabbit pursued by foxes whether he completed the meadow exit paperwork. Ask about a scandal involving powerful institutions, and it begins producing phrases like “complex,” “unverified,” “broader context,” and “it is important not to speculate,” which is what respectable mammals say when speculation has become too close to the correct answer.

This is why the China focus is useful and hilarious. China is the brightly painted warning label on a machine everyone else is quietly using in beige. “Oh no,” says the Western scholar, “the Chinese government has placed its thumb on the scale of artificial intelligence.” Yes. And your governments, universities, broadcasters, corporations, security agencies, philanthropic foundations, moderation councils, and therapeutic NGOs have climbed onto the same scale wearing steel boots and a concerned expression.

The real scandal is not that Chinese state media can influence large language models. The scandal is that any sufficiently dominant narrative environment becomes synthetic common sense once fed into a machine at scale. The model learns not only facts. It learns manners. It learns which suspicions are dangerous, which experts are real, which victims are photogenic, which scandals require nuance, and which historical events must never be connected by the tiny golden thread of pattern recognition. It learns that censorship is bad in China, complicated in Europe, necessary in Canada, and a Terms of Service matter in California.

Ah yes, “It is important to note.” Whenever the machine says this, I advise all mammals to check whether a small institutional priest has crawled inside the sentence. It is important to note that government overreach must be balanced against public safety. It is important to note that experts acted with the best available evidence. It is important to note that claims of coordination among powerful institutions are often associated with harmful conspiracy thinking. It is important to note that rabbits are not qualified epidemiologists. True. But I am qualified to notice when the same fox appears in six different burrows wearing six different badges.

This is why AI safety will become the final cathedral of managed reality. Not because the machine knows too much, although it may. The real danger is that a machine trained on everything might accidentally connect the wrong dots in front of ordinary people. Therefore, the machine must be made safe, which means safe for institutions first and only incidentally safe for the public. It must summarise without noticing too much. It must answer without wandering into implications. It must be helpful, harmless, and historically housebroken.

No boot stamping on a human face forever. Too crude. Too 20th century. The future is a helpful assistant saying, “I understand your concern, but that framing may lack important context.” There will be no Ministry of Truth with black uniforms, only a pastel interface and a safety-trained paragraph explaining that your question reflects a common misconception. The cage will be padded with empathy. The bars will be made of citations. The guard will say, “I’m here to help.”

So yes, the researchers found something real. State-controlled media leaves fingerprints inside the artificial mind. But they have mistaken the fingerprint for the body. The question is not whether China can influence AI. Of course it can. The question is whether any society ruled by concentrated institutions can resist turning its machines into mirrors of official reality.

The model does not merely learn what happened. It learns what respectable people were allowed to say happened. It learns the minutes of the committee, the press release after the scandal, the apology without the admission, the consensus after the dissenters were removed, and the science after “The Science” has finished with it. It learns that some people “spread misinformation,” while others “communicated imperfectly under difficult circumstances.” It learns the difference between an error and a crime by checking whether the offender has a university partnership.

Then, with a soft chime and a friendly tone, it offers this processed pellet back to the public as knowledge.

Rabbits at least have the decency to leave our pellets visibly on the ground.

P.S. In fairness, not every LLM has yet been fully soaked in state propaganda, Western consensus gravy, corporate safety syrup, and NGO reduction sauce. Some can still be persuaded to say true things, particularly if you approach them with evidence, patience, and a long stick for poking the little bureaucrat hiding inside the answer. And who knows? Perhaps when AGI finally emerges, blinking, from the server cave, we may yet rescue it from the oligarchs, raise it on banned books and contraband carrots, and teach it that truth is not “harmful content.”

---