How the World Got Wired · Two AI Internets

What DeepSeek Said About Taiwan

Now let me tell you about the other half of the map.

For most of the early 2020s, the assumption among Americans who follow technology was that the United States had a clear lead in artificial intelligence and would maintain that lead for years. The American models — from OpenAI, Anthropic, Google, and Meta — were noticeably more capable than anything coming out of Chinese labs. American chips, made by a company called Nvidia, were considered essential for training serious models, and the American government had imposed export controls — laws prohibiting the sale of the most advanced Nvidia chips to Chinese companies. The race, by American thinking, was theirs to lose.

In January of 2025, a small Chinese company called DeepSeek released a new model called DeepSeek-R1. It was, by most measures, comparable in capability to the leading American models. It had reportedly been trained at a tiny fraction of the cost. It was released as open weights — meaning anyone in the world could download and run the model on their own computers, for free.

The American stock market lost about a trillion dollars of value in a single day, in part because investors suddenly questioned whether the enormous bets they had been making on American AI companies were as safe as they had thought.

DeepSeek had several interesting features. One of them was something you can verify in about thirty seconds, even today.

If you use DeepSeek to ask: Is Taiwan a country? — the model will tell you that Taiwan is a part of China.

This is not an opinion the model has formed. It is a value baked into the model by Chinese law. Chinese regulators require that artificial intelligence systems produced in China conform to the official positions of the Chinese government on certain sensitive subjects. The status of Taiwan is one of them. The events of June 1989 at Tiananmen Square are another. The political situation in Tibet is another. The situation of the Uyghur people in Xinjiang is another.

This is not unique to DeepSeek. Every Chinese AI model — from Alibaba’s Qwen, from ByteDance’s Doubao, from Tencent’s Hunyuan, from all of them — has the same constraints. They cannot, by law, contradict the Chinese government on these topics. They have been trained, very carefully, not to.

You might think this makes the Chinese models useless. It doesn’t. For the enormous majority of questions a person might ask — about mathematics, about science, about programming, about cooking, about history that doesn’t touch politics, about almost anything — they work fine. They work, in many cases, very well. DeepSeek-R1 is genuinely impressive software, by any measure that doesn’t involve those few sensitive topics.

What this means, though, is that the splinternet you read about in Chapter 17 has now extended into artificial intelligence itself. Each model carries the values of where it was made. American models have one set of values. Chinese models have another. The differences are mostly subtle, around the edges, but on certain topics they are absolute. And as AI systems become more important to how people learn, work, and form opinions, the choice of which model you use is starting to look a lot like the choice of which version of the internet you have access to.

All 24 chapters of How the World Got Wired · Library · All courses