LEILA FADEL, HOST:
President Trump wants a preview of the newest AI models. An executive order asks companies to submit their products to the government for review up to 30 days before their release date. Last month, the president delayed the order, saying he worried the U.S. could give up its lead in artificial intelligence. Our next guest worked on AI policy in the Biden administration as acting director of the White House Office of Science and Technology Policy. Alondra Nelson is now a professor at the Institute for Advanced Study. Good morning, and welcome to the program.
ALONDRA NELSON: Good morning, Leila. Thanks for having me.
FADEL: So how significant is this order from a White House that previously was very hands off towards AI?
NELSON: Yeah. Well, I mean, it is a move in the right direction, and I think it represents something of a sea change. I mean, from the campaign up until Vice President Vance's first international speech, you know, the - this administration has been saying, we will do nothing, government will do nothing to sort of intervene or trying to lead or shape the trajectory of artificial intelligence development in the United States. And indeed, Vice President Vance, at this international speech in Paris, said something like, you know, we shouldn't be hand-wringing around safety. Like, we shouldn't have to, you know, have regard for issues of safety and security at all. So it certainly represents a sea change in some regard, and it's an important first step, I think.
FADEL: It says asks. What does that mean? I mean, is this just a suggestion?
NELSON: Yeah. So yeah, we - I think we want to be clear that this is not regulation. I mean, it's an executive order, and it does a couple of important things. One is that it directs the federal agencies, which executive orders usually have purview over, to do a few things. And one of those things is to work with industry to get a 30-day preview, as you said, of advanced models. So this is different from what - the executive order that was going to be released last month that was not released. That had asked for a 90-day window to - of preview. So this is a 30-day window, but it's all voluntary. So, you know, you're - we're relying on the industry to sort of give us the models, show the government the most dangerous models, the models they think that are most risky. And this is all happening in the context of this AI race - right? - for...
FADEL: Yeah.
NELSON: ...Markets for - you know, we - there's a new IPO from Anthropic that's being - that's just been announced. So there's a lot of competition here. But - and there's going to be classified benchmarks. So there'll be some information about how these models are sort of testing on various kinds of catastrophic risk. Unfortunately, none of this will be made available to the public, even at a high level. So we won't even know kind of the state of things as an American public.
FADEL: You mentioned safety. You mentioned security. I mean, Trump's order says, quote, "advanced AI capabilities make our nation stronger but also introduce new national security considerations." What AI capabilities are drawing the attention of the Trump White House?
NELSON: Well, it was - we've known for a very long time and industry leaders, academic researchers have told us that there had been cybersecurity risks potentially with AI models, in addition to risks of sort of chemical and biological warfare and other kinds of risks when you can automate sort of science and other kinds of processes. So that had been known.
I think what triggered this was really in April when Anthropic released Claude Mythos, this powerful model that surely had capabilities that could reveal very quickly vulnerabilities in cybersecurity. So companies that thought that they had shored up their systems, you know, that had shored up critical infrastructure - so this is that architecture that really is the backbone of all of our lives, everything from energy to banking to hospitals and more - that we had shored that up. And with Claude Mythos, here you - we're being told you had a system that could very rapidly find dozens if not hundreds of vulnerabilities in the systems. So to release that out into the world - which Anthropic, to its credit, did not - would've been incredibly dangerous and just would've created sort of chaos in the world. So I think the challenge with this executive order is that Anthropic acted independently to sort of roll out slowly and make people aware that this model was coming out. It's not clear that government requiring a preview will be any different from companies taking that action.
FADEL: The order says the Biden administration put bureaucratic constraints on AI. As someone who worked in the Biden administration, how would you describe how this is different from the policies you worked on?
NELSON: So there was an exact same sort of preview that existed in an entity called the U.S. AI Safety Institute. And so this was something that the administration - the Biden administration identified in 2023 was necessary. The government has to have some role.
FADEL: We'll have to leave it there, unfortunately. I'm so sorry. Alondra Nelson worked on AI in the Biden administration and is a professor at the Institute for Advanced Study. Thank you.
NELSON: Thanks, Leila. 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.
