SACHA PFEIFFER, HOST:
The new dirty word in the world of artificial intelligence is distillation. It means using someone else's AI to train your AI. The Trump administration has accused a Chinese startup of using this process to steal from the U.S. AI company Anthropic, and that has sparked a big debate about whether this process is legal or ethical and about the risks it creates. Here to boil it down for us is NPR tech correspondent John Ruwitch. Hi, John.
JOHN RUWITCH, BYLINE: Hey, Sacha.
PFEIFFER: Tell us more about this accusation.
RUWITCH: Yeah, the Beijing-based startup Moonshot AI launched a new model this month called Kimi K3. It's quite good. Then, in a post on X, though, President Trump's science and technology adviser Michael Kratsios said K3 was developed by distilling an Anthropic model. I reached out to both companies. Moonshot did not respond, but Chinese media reports that the company has denied the claims. And Anthropic hasn't pointed a finger at Moonshot over this particular model, but it says distillation by Chinese companies is happening and poses a national security risk because these models are getting better and could be dangerous when it comes to things like cyber- or biosecurity.
PFEIFFER: John, the word distillation. I imagine some people think of it more as a chemistry lab word or an English paper word. Tell us...
RUWITCH: Right.
PFEIFFER: ...More about what it means in this context.
RUWITCH: Yeah, so basically, distillation is when a company trains a new AI model by running tons of prompts through a better, existing AI model. It can see what good answers look like and sometimes how the AI got those answers. It's a shortcut, basically, that takes advantage of the work of others. The problem is that companies like Anthropic spend billions and billions of dollars creating their cutting-edge AI models. They're offering top talent in the industry elite-professional-athlete-level pay packages. They're pouring money into data centers. China has great engineers, but you need computational muscle to build AI, and the U.S. blocks the sale to China of the best microchips, so they're coming up with other ways to supercharge their AI.
PFEIFFER: You know, it seems natural that companies in China - or anywhere, really - might want to build off the work of U.S. companies by using distillation, but is it actually illegal? Is it legal?
RUWITCH: Yeah, this is all so new that none of this has been tested in court, but legal experts do say there are some potential ways companies could try to defend their models with the law. You know, one possibility is they could argue distillation is copyright infringement. Generally speaking, though, AI-created answers have not been considered copyrightable, so that avenue isn't a sure bet, but, Sacha, it would be interesting.
PFEIFFER: Yes. In fact, it would certainly be ironic for AI companies to accuse others of violating copyright rules, given the vacuuming up of so many publications they've done over the years.
RUWITCH: Yeah. I mean, big American AI companies have faced a lot of lawsuits over that from book publishers, newspapers, others. They claim that they stole copyrighted published works to train their own AI models. Here's Anupam Chander. He's an expert in this new area of tech at the Georgetown Law School.
ANUPAM CHANDER: After all, the AI companies believe, and they have argued, that learning from others is a perfectly fair use of other people's copyrighted work.
RUWITCH: Now, fair use is when material can be used without permission for research, teaching or commentary. AI companies could also potentially argue that distillation violates the terms that a user agrees to when they sign up to use an AI model or that companies doing the distilling are stealing trade secrets. It's unclear if any of this would stand up in court, and even if it does, if we're talking about companies, you know, in China - going after companies in China - there are questions of enforcement.
PFEIFFER: And as we said, the Trump administration objects to this. Tell us more about what they're saying.
RUWITCH: Yeah, the administration accused Moonshot of distilling from Anthropic. And then Treasury Secretary Scott Bessent last week said the administration had evidence that many Chinese AI models were doing the same. He said on X the administration was looking into it and could impose sanctions. That got Beijing's attention. On Monday, China's Commerce Ministry warned that it could retaliate if the U.S. takes steps that violate China's rights and interests.
PFEIFFER: That is NPR tech correspondent John Ruwitch. Thank you, John.
RUWITCH: You bet.
PFEIFFER: And a note that Anthropic is a financial supporter of NPR. Transcript provided by NPR, Copyright NPR.
NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.
