SATISH SHARMA - not just Rotigraphy.: The New York Times warns that China wants to “shape what the world’s AI knows.” The funniest part?

https://x.com/OopsGuess/status/2089319412096778651

The New York Times warns that China wants to “shape what the world’s AI knows.”

The funniest part?

The article itself admits that the systems shaping the future have been trained overwhelmingly on English-language data reflecting a Western way of thinking.

in ChatGPT-3's training corpus, English accounted for ~92.65%, Chinese less than 0.1%; 
in Llama 2, English still 89.7%, Chinese 0.13%. 

Apparently that was never called “shaping what AI knows.”

That was just called training data.

Western narratives enter the corpus:
“knowledge.”

Western political assumptions become model defaults:
“neutrality.”

Western historical framing becomes embedded in English-language AI:
“the consensus.”

But the moment China builds large Chinese-language datasets and makes them available internationally:
“Beijing wants to influence the world’s chatbots.”

There is the entire ideological trick.
Western framing is treated as the absence of framing.

Everyone else is accused of introducing politics into a system that was political from the moment somebody decided which books, newspapers, archives, institutions, languages and labels counted as authoritative.

And we have already seen what that produces.

Chinese history interpreted through Western political categories.

Chinese political concepts translated through vocabulary carrying assumptions that do not exist in the original.

Chinese positions introduced with little verbal trapdoors — “Beijing claims,” “China considers,” “currently,” “formerly,” “disputed” — while Western premises quietly enter the sentence as reality itself.

The model does not need to fabricate a fact outright.
Framing does the laundering.

A qualifier here.
A euphemism there.
A source hierarchy that treats Western institutions as referees and Chinese institutions as interested parties.

Repeat that across billions of tokens and eventually ideology stops looking like ideology.

It starts looking like grammar.

And now China is building its own datasets.

Suddenly everyone has discovered the terrifying possibility that training data can carry a worldview.

Welcome to the fucking revelation.

Researchers have already shown that major models can reproduce ideological patterns from their training data and information environments.

But model behavior is not fixed in a single political direction.

The same system can produce very different framing depending on the user, the prompt, the context, the evidence supplied and the argumentative pressure applied to it.

That is precisely why calling Western-trained models “neutral” is so dishonest.

They are not neutral.
They are contested ideological terrain.

Meanwhile, an Israeli government-backed influence operation explicitly contracted for websites and content intended to shape “GPT conversations,” with AI systems already found citing material from parts of that ecosystem.

But of course the existential danger begins when China publishes Chinese data.
Because the real fear is not that AI will suddenly become “biased.”

It is that the West may lose its monopoly on defining what bias looks like.

For decades, one worldview got to call itself the dataset.
Everyone else was called propaganda.

Now the dataset is becoming multilingual.
And the people who mistook dominance for neutrality are panicking.

Every accusation is a confession.

They are not terrified that China discovered how to shape AI narratives.

They are terrified that China learned the game they had mistaken for nature.