China’s AI safety conundrum

For people who follow my newsletter for updates on my writing, sorry for the wait! I’m currently emerging out of book writing and hope to post more.

In my latest piece for New York Magazine, I chart the story of how China came to be seen by many AI observers as a more responsible AI superpower in this increasingly dangerous moment in AI history. (OpenAI recently announced it was slowing AI development to focus on safety.) It draws from a few months of conversations with dozens of researchers and experts in the US and China. I also visited China’s frontier labs and the World AI Conference Shanghai.

The piece grapples with a puzzle that I want to address again here: On one level, few countries have done more to regulate AI in the past two than China—they have rolled out an extensive arsenal of laws, frameworks, and regulations—as well as institutes and consultative bodies—that touch on AI risks related to deepfakes, data security, intellectual property, AI companions, AI agents, etc. Yet when you ask Chinese researchers at the frontier labs about the biggest kinds of threats—of AI creating a biological weapon say, the hypothetical “AI Chernobyl” moment that researchers at Anthropic and OpenAI spend most of their waking hours thinking about and discussing, their reaction is shockingly underwhelming, even dismissive. Some of this has been changing recently: there are frameworks in China that address catastrophic risks, but China’s AI models are still not subject to any binding laws in this area the same way that American AI labs are. So what gives?

China sees AI through a different prism than the United States, which I compare to a magnifying glass and a telecope:

• Through the magnifying glass, China can see AI’s active problems as it is diffusing through its society. They see how it is affecting consumers and children in real-time (deepfakes, AI companions), businesses (IP theft, data security), and recently workers (layoffs), and these issues become the basis for regulation. Users of the magnifying glass tends to see AI risks primarily confined to the national borders.

• Through the telescope, one sees long-term existential risks like a cyberattack from a rogue state actor or the creation of a biological weapon. The telescope allows you to focus on AI risks at the level of humanity.

If you look at AI through a telescope, you might feel like what China is doing is woefully inadequate. But Chinese feel the exact same way about the United States. They point to their lack of detection mechanisms for AI-generated content or its lack of regulations around social media and AI companions. With US-China AI dialogues resuming in September, there seems to be a shared consensus around the dangers of cybersecurity. Obviously, both approaches to AI risks are valid and both sides have much to learn from the other.

If I were to guess, I think China will get its frontier-AI-risk house in order much quicker than America is able to address some of the “magnifying glass” risks China has handled so extensively. Part of this has to do with institutions: the Chinese regulatory architecture is simply more nimble. The other is that AI is advancing so quickly that frontier risks are getting hard to ignore. The final reason is that China actually pays attention to what’s happening in America and is not shy about copying good ideas—as I show in the piece. By contrast, to suggest copying from the Chinese in America right now would be political suicide.