Why banning open-source AI would backfire

Washington is moving toward tighter control of frontier AI: an executive order to review models, draft legislation in Congress, talk of the government taking equity stakes in the big labs, and new limits on which foreign nationals can use the most advanced Anthropic models. In a piece they say traditional outlets turned down, Nathan Lambert and Kevin Xu argue that extending any of this to open-source AI would be a serious mistake. Their core point is simple: open source is just a method for sharing and building software in the open, and it has a long record of being safe and economically useful.

The numbers they cite are hard to wave away. More than 90 percent of the world's software is built on open source, which they estimate has generated over 8 trillion dollars in value. Linux runs more than 90 percent of cloud infrastructure, Android opened up smartphone competition, and Meta built its early products on open tools. On safety, they lean on the old observation that with enough people looking, bugs surface fast, and they note that open models run on your own machines do not ship your data anywhere.

The sharpest part of the argument is about China. Lambert and Xu say that cracking down on open weights to slow Chinese labs would do the opposite of what is intended. It would chill American work while the rest of the world, deprived of competitive open models from the United States, simply adopts Chinese ones instead.

Why it matters

If you build on open weights, the rules being drafted now could decide whether you keep that option. The authors are telling builders and policymakers to separate the real concern, frontier capability, from the distribution method, and to watch the coming legislation closely before it quietly closes a door.