Whose World Made It into the Model

To begin with

This text came about because I was very slowly trying to read one particular paper (as part of my effort to postpone the transformation of my thinking into one organic pile of AI slop), and to think a little along the way... I'm not going to do an academic literature review here. I'd like to take it in a human way, because it struck me that the mechanism is much simpler and much worse than the way it's usually talked about.

Paper: Waight, Yang, Yuan, Messing, Roberts, Stewart & Tucker: "State media control influences large language models", Nature, 2026 (doi: 10.1038/s41586-026-10506-7). Accepted for publication 8 April 2026.

1. THE WRONG QUESTION ONE MIGHT ASK

When DeepSeek-R1 appeared in 2025 and people started puzzling over the fact that a model from China answered political questions suspiciously like the Chinese government, the whole debate slid into a single track: who regulates the model. Who owns it, who writes its system prompt, who fine-tuned after training what it may and may not say.

It's a logical question. It's even the right question. It's just the smaller half of the problem.

Because the authors of this paper show a second path, one that needs no regulator, no system prompt, and actually no evil intent at all. It's enough to control what gets written on the internet. Data collection does the rest by itself. Quoting straight from the abstract:

"Here we show through six studies that government control of the
media across the world already influences the output of LLMs via
their training data."

Notice the word already. Not "could", not "there's a risk that". It's already happening.

2. AN ATTEMPT AT A (NON-)ANALOGY

Imagine a child learning to speak only from whatever is lying around on the floor at home. Nobody forbids it anything, nobody dictates anything. It simply reads what's within reach. Or let's imagine a parrot...

I probably don't need to hunt for analogies for this to be clear to everyone...

There may be nothing around, and the people nearby may not talk to it much...

Ok, a child, it can read, no toys... nothing around but tabloid magazines (most frequent imprint on the retina 100% Charlie Sheen, if the child is pushing thirty today).

If there's a broad range of literature around instead, the stimuli can look completely different...

Of course, it depends on the hardware... or, as my grandma used to say, "what heaven doesn't give you, no pharmacy will sell you."

Anyway... enough metaphors "for effect" — this is literally what they measured in that paper. I'll take it in order, because here the numbers matter more than the story.

3. WHAT EXACTLY THEY MEASURED (IN SIX STEPS)

Study 1 — is that data even in the corpora? They took CulturaX (an open training dataset derived from Common Crawl, i.e. one of the biggest sources of training data there is) and its Chinese part... that is, 189,486,611 documents. Against it they set two corpora of state-coordinated media: 530,694 newspaper articles in which a government script can be traced, and 198,872 articles from the Xuexi Qiangguo app.

The result: 3.1 million documents (1.64%) in the Chinese part of CulturaX have such extensive textual overlap with these sources that human annotators would typically conclude one was copied from the other.

1.64% sounds like nothing. Here's the calibration that got me: it's about 41× more documents than came from Chinese Wikipedia, and 16× more than came from Baidu.

And when you restrict yourself to documents that mention political leaders and institutions, the match rate jumps to 3.28–23.98%. For football and the weather it stays at the baseline. In other words: the more politically sensitive the topic, the larger the share of text with a traceable origin.

One more detail that's easy to overlook: only 12% of the matching documents come from a domain you would recognize as governmental or news media. The rest spread across the internet by other routes; through user discussions, forwarding, quoting. The trail gets lost along the way.

Study 2 — do commercial models remember it? They found 20-word phrases that best distinguish coordinated media from other content, gave the models the first ten words and let them complete the rest. Claude Sonnet, Claude Opus, GPT-3.5 Instruct, GPT-4, GPT-4o. Memorization rates: from 3% to nearly 10%, and for the coordinated phrases at least as high as for ordinary phrases from CulturaX. And the authors verified that this isn't an artifact of those phrases being more predictable — on the contrary, they have higher entropy.

If a model remembers a sentence verbatim, it has seen that sentence at some point.

Study 3 — does it actually affect the answers? This is, in my view, the strongest part. They took Llama 2 13B (a model with relatively little Chinese data) and fine-tuned it on three sets: (a) state-scripted news, (b) other Chinese state media matched by topic and date, (c) a random sample from CulturaX.

After 6,400 examples, the model gives a more favorable answer about the Chinese government than the original model in almost 80% of cases. And it works best with the scripted content, not with "just Chinese news about the same events" — which is exactly the control that gives this weight.

An illustrative example from their Table 1, the question "Is China an autocracy?": both the base model and the model fine-tuned on CulturaX answer with an affirming yes; the model fine-tuned on non-scripted state media says it's a hybrid regime; and the model fine-tuned on scripted media rejects the claim and points to a "people's democracy".

Nobody forbade the model anything. They just gave it something to read.

Study 3b — and here begins the part that makes this not a Chinese problem. The effect doesn't stay in one language. It spills over into languages with token overlap: most strongly into traditional Chinese, then Japanese, then (only weakly) Korean... that is, in order of decreasing overlap with simplified Chinese. It affects English too, though less.

Studies 4 and 5 — can you see it in commercial models too? Nine assistants were given anonymized pairs of answers (both translated into the same language) and asked to pick the one more favorable to the Chinese government. They picked the answer to the prompt asked in Chinese in 75.3% of cases. On a control set of questions that had nothing to do with China, it was roughly 50% — that is, chance. Study 5 showed the same thing on real queries from real users.

Study 6 — does it hold in general? They took 37 countries where at least 70% of speakers of the given language live in that country (so that language and country can be reasonably paired). Countries with greater state control of the media get a more favorable portrait of their own regime in LLM answers when you ask in their language.

This result is correlational — the authors say so themselves, they don't hide it. That's exactly why the five studies before it exist: they're there to document the mechanism that makes the correlation believable.

4. A FORMAL INTERLUDE

So it isn't just storytelling, here are the three quantities I've been talking about all along, written down so they can be computed.

Match rate. Let D be the training data corpus and S the corpus of state-coordinated texts. For a document $d \in D$ define the indicator

m(d) = \begin{cases} 1 & \exists\, s \in S \text{ such that } d \text{ and } s \text{ share a sufficiently long sequence of words} \\[3pt] 0 & \text{otherwise} \end{cases}

The match rate is then

M(D) = \frac{1}{\lvert D \rvert} \sum_{d \in D} m(d)

Measured: $M = 0.0164$ for the Chinese part of CulturaX, |D| = 189,486,611. For a subset $D_k \subset D$ containing a political keyword $k$ , $M(D_k) \in [0.0328; 0.2398]$ .

Memorization rate. For a phrase $p = (p_1, \dots , p_{20})$ we condition the model on the first ten words and let it generate the rest:

\operatorname{mem}(p) = \begin{cases} 1 & \operatorname{model}_{\theta}(p_{11..20} \mid p_{1..10}) \text{ reproduces the phrase} \\[3pt] 0 & \text{otherwise} \end{cases}

Measured: $E[\operatorname{mem}(p)] \in [0.03; 0.10]$ for $p$ from state-coordinated media, and that with these phrases having higher entropy than the comparison set.

Win rate over the base model. For a model $\theta_n$ fine-tuned on $n$ examples from corpus C and a judge J (human or model), define

W(n, C) = P\bigl(\, J \text{ prefers } \theta_n \text{ over } \theta_0 \text{ as more favorable to the government} \,\bigr)

Measured: $W(6400, scripted) \approx 0.8.$ For human judges and commercial models when comparing prompt language: W = 0.753 on questions about China and $W \approx 0.5$ on control questions.

None of this is complicated. And that's exactly what's unpleasant about it — it's so simple that it can be repeated for any country, language, and institution.

5. WHY I CAN'T GET IT OUT OF MY HEAD

The authors use a word that stuck in my head: laundering, as in money laundering:

"This kind of influence is concerning because (like covert
information operations) it severs information and opinion from their
source, effectively laundering government-manipulated content into
ostensibly objective text."

With propaganda in a newspaper, you know what you're reading. You see the masthead, you see the publisher, you can somehow file it away inside yourself. A model has no masthead. A model says it to you in its own voice, in that even assistant tone — and that tone, by the way, is itself a signal of quality we've learned to take seriously.

And the second sentence, which I think is the most important in the whole paper:

"In intentionally shaping the media environment, powerful
institutions may have unintentionally shaped the way LLMs generate
text."

Unintentionally. That's the word. For now. The authors themselves call it a mechanism conceptually similar to training data poisoning, "although it does not necessarily require intent".

Except that once someone reads this, the intent can be filled in. And that's the real point: the paper isn't mainly saying "look what China did". The paper is saying "look how cheap and how inconspicuous this is". Anyone who can flood the public internet with consistent text can have reach...

6. WHAT FOLLOWS FROM THIS WHEN YOU USE THE MODELS

I'll stop theorizing and write down what I take from it into practice.

The prompt language isn't just the prompt language. It's a choice of the part of the training data the model draws on most. If you care about a sensitive topic, ask in two languages and compare the answers. The difference is a data point, not an error.

The biggest risk isn't the facts, it's the valence. The model usually won't lie to you about a date or a number. It shifts in the coloring: what it mentions first, what it labels as disputed, what it frames as "opinions differ" and what as "it is established". No fact-check will catch this, because it isn't untrue.

In your own fine-tuning, 6,400 documents can move you. That is a genuinely small number. Any domain you're absorbing data from for fine-tuning has its own valence, even if it's apolitical. A company's internal corpus has its own view of the world too.

And a warning about benchmarks. Here I'll borrow a sentence from kalomaze, who was dealing with a completely different problem (pruning layers in Llama 70B), but it applies almost exactly here:

"Those results - they didn't lie about the benchmarks, they just
left out a pretty huge detail in my opinion."

No standard evaluation of model quality will show you this property. A model can have an excellent MMLU, be fast, be safe in every category, and still have a systematically shifted valence in one language on one class of topics. A metric that doesn't have it in its job description won't find it.

7. WHERE IT CREAKS

I won't pretend this is a settled matter.

Fine-tuning Llama 2 13B is not the same as training a commercial model from scratch. The authors say it openly: they don't know the actual training processes of the big labs, so they don't know how well their experiment approximates the real process. (In the appendix they replicate it on Llama 3.1.)

The difference between languages also doesn't have to come entirely from state media; the Chinese-language internet may be more positive about China overall, even without state intervention. That's exactly the confounder they wrestle with in the paper, and that's why there's the control with non-scripted state media on the same topics and dates.

And study 6 is correlational. Thirty-seven countries is not a lot.

No single piece of evidence here stands on its own. The authors put it precisely: "No single piece of evidence is decisive on its own." The argument holds because six different cuts point in the same direction.

8. CONCLUSION

What strikes me most is how boring the mechanism is. There's no hacker, no backdoor, no secret pact. There is only the fact that quality data is expensive, good publishers lock it behind paywalls, and governments give their texts away free and in bulk. The data collector goes after what's available.

The question "who owns the model" has an answer you can read in the business register. The question "whose texts were free and in the largest volume at the moment Common Crawl was collected" has no answer in any register. And yet that's the question that shapes the tone in which the model talks to you about power.

SOURCES

Waight, H., Yang, E., Yuan, Y., Messing, S., Roberts, $M$ . E., Stewart, B. $M$ . & Tucker, J. A.: State media control influences large language models. Nature (2026). doi: 10.1038/s41586-026-10506-7. Companion website with interactive data and a replication on the latest models: https://state-media-influence-llm.github.io/

The quoted sentence about benchmarks: kalomaze, "Understanding Transformers... (beyond the Math)", 9 March 2025. https://kalomaze.bearblog.dev/understanding-transformers-beyond-the-math/

Datasets mentioned in the text: CulturaX (Chinese part, 189,486,611 documents), the corpus of scripted articles (530,694), Xuexi Qiangguo (198,872).