Hey everybody, this is Andrew Egger with The Bulwark. It has been an absolutely insane weekend at the Pentagon. Obviously, the breaking news right now is that we may or may not now be at war with Iran. Guess we're going to find out a lot more about that. We did a live stream about that earlier this morning. We're here to talk about something completely different happening at the Pentagon. A very crazy story. We thought it was going to be the craziest story out of the military certainly this week until it was superseded by events, but maybe a story with just as important long-term repercussions. We're going to find out. We're talking about the Department of Defense's spat, which is now sort of blown up into an all-out war against the AI company Anthropic, perhaps better known to you guys out there as the producer of the chatbot Claude. They have gone basically nuclear in this fight with this AI producer that was formerly a top defense contractor for the DoD. And I have just been following this story in my newsletter, Morning Shots, for the last week or so, but it's obviously a much longer story that's just kind of coming to a head now. So, I'm very glad to be joined by Hayden Field to talk all through this. She is a senior AI reporter for The Verge. She knows all this stuff backwards and forwards. So, Hayden, thanks for coming on to talk about this story today.
Yeah, thanks so much for having me. It's been a crazy week for sure.
Oh my lord. Okay, so let's just start. Maybe you can give us just kind of the pre-existing, like two weeks ago we were at a very different status quo. There were like inklings that this fight was happening, but can you just kind of walk us through how things have fallen apart between Pete Hegseth and Anthropic over the last 10 days or so?
Absolutely. So it all really starts back on January 9th when Hegseth sent a memo saying, you know what, I want to renegotiate all of our existing AI contracts to be for any lawful use. So before, you know, AI companies could and did put their own terms into these contracts, like you can't use it for this. If you're going to use it for this, you have to do X, Y, and Z. He was saying, "No, I want to take all that out and just remove all the barriers." So, you know, obviously that kicked off a bunch of negotiations. Now, apparently, my sources told me they were in pretty good faith for a while. Then, you know, about 10 days ago, like you said, things seemed to turn pretty ugly. I think that negotiations had stalled a bit. Emil Michael had been tweeting a lot of stuff. There were a lot of public social media posts kind of being traded back and forth. There were public statements being traded back and forth. Lots of insults. Things were getting more inflammatory. Anthropic was standing by its guns where it said, "Hey, you know, we're not okay with domestic mass surveillance and we're not okay with lethal autonomous weapons," which basically means AI being used to kill people with no human oversight. So, those were their two red lines. They were like, "We're not budging on these two things. That was already in our existing contract with you guys. So, let's just keep it the same or let's try to find a compromise here." It seemed that the Pentagon really wanted any lawful use and they were not okay with any, you know, type of exception to that. So, yeah, things have gotten a bit uglier. You know, I've been covering this. Yeah, every day something's different and on Friday, honestly, every 30 minutes something changed. You know, for a while we thought that they were going to reach a deal, a compromise of some sort. You know, even when I spoke with people at Anthropic that day, like pretty late in the afternoon, they thought things might have gone differently. But, you know, 5:00 p.m. rolled around, the deadline that they had been offered to either acquiesce or else. And then a social media post went out saying that they were going to be labeled a supply chain risk, which it's strange because that's something that usually they would never label a US company. It's something that usually it's reserved for like foreign adversary companies or ones that might have some type of like cybersecurity risk. Usually companies in China, for example, are on that list. But yeah, never a US company that we know of so far. So it was an interesting thing. It raised a lot of red flags. I heard from both parties. People were worried about, you know, if you disagree with the Trump administration, could you just be labeled a supply chain risk just randomly? You know, those were the kind of questions that were being asked. And yeah, basically what that means, we're not 100% sure on what the exact granular parts of this would entail, but essentially it looks like if you work with Anthropic and you're, you know, a defense contractor or something, you need to provide a version of your services without any involvement with Anthropic to the Pentagon or the DoD. So that's going to impact Anthropic's business a lot. Not their consumer side, but they do a lot of enterprise business too, a lot of military business. So, yeah, it's definitely going to be an interesting couple days. I'm sure we'll both be working all weekend on the updates.
It is hard. It has been hard for me really to process just how surreal this whole story has been because one of the defining characteristics of Anthropic amid sort of like the leading AI labs as far as their relationship with the government is concerned just in the last couple of years has been that they have been sort of at the forefront of like leading the charge to integrate with the Department of Defense to say, yeah, we think it's great for our models to be used for these national security purposes. You know, we think that's very pro-democracy for, you know, the American government to be able to deploy these things, you know, against authoritarian regimes around the world. They were previously the only lab that had a contract to deploy these AI models in classified settings, period. Like, xAI didn't have one. OpenAI didn't have one. Google didn't have one. And, you know, these things had gotten pretty integrated in the Pentagon's, you know, war planning and the work they're doing behind the scenes there. They were reportedly used, for instance, when the Pentagon went in and got Nicolas Maduro, the dictator of Venezuela, a few months ago now. And so it's because of how tightly they have been integrated so far, like because it seemed like this was such an edge case dispute, right? Like about hypothetical potential future lethal autonomous weapons. I mean, maybe correct me if I'm wrong here. My read on this, and I think a lot of people that I was talking to, a lot of my sources' read on this, was that Hegseth's threats of labeling them a supply chain risk over all of this were so sort of over-the-top and histrionic that a lot of people just thought, well, this is just sort of the biggest stick he has to sort of shake at Anthropic right now to try to bring them around. And that if Anthropic still stuck to these red lines, that Hegseth might tear up the government contract. That would not be very good for Anthropic. Might move to some of these other models. But that almost wasn't even in people's real sense of a genuine realm of possibility. Am I crazy about that? How was Anthropic feeling about that at the time?
No, that's exactly what I was hearing from all my sources. They thought that it was, you know, kind of just like everyone was playing chicken. You know, everyone's sticking to their guns saying, "Oh, we're not moving at all. We're not moving at all." And that by the time the deadline rolled around at 5:00 p.m. on Friday, that everything would change and there would be some type of compromise of some sort, or at least, like you said, maybe he would tear up their government contract that wouldn't put them out of business. I mean, you know, it's a lot of money, but it's not... they have a lot of other business that makes them just as much money. So, right, that's really what people thought was going to happen. Now, this was crazy because I had also been hearing that xAI and OpenAI had also signed the terms already. No problem. That's what some reports were saying. Now, when things got really intense yesterday afternoon and, you know, all the negotiations were coming to a head and we didn't know what was going to happen, OpenAI CEO Sam Altman apparently sent out a memo internally and said, "Hey, I'm working on a deal here. Stay tuned. We have the same red lines as Anthropic." That's what he said. Now, later last night, he said that they did end up coming to a new deal with the Pentagon and saying, you know, we got the same terms. Basically, he's implying that Anthropic was fighting for, but we also got to keep our contract. You know, we're going to try to make sure that the DoD can kind of give this same deal to all the other labs, and we think they should sign. Now, what he's kind of implying there, and the way a lot of people read it was that I think he's trying to get the clout of saying, "Hey, like we got the same terms just because we were playing nice and now we're going to, you know, be the heroes and kind of offer, like, encourage the DoD to give this same deal to all the other labs." But if you read the fine print there on his statement, it looks like he signed maybe a lesser deal than Anthropic was fighting for. Maybe the domestic mass surveillance or the lethal autonomous weapons. Things seem to be worded a little bit differently in his statement. So, I'm working on that now, but it seems that, yeah, the deals were maybe a little bit different. We just don't know how it changed the wording and the definitions of these things and what exactly they acceded to here that Anthropic was kind of fighting against yesterday.
Yeah, I'm glad you said that because I'll just say I have zero sources at OpenAI. I cannot call up anybody at OpenAI and get their kind of read on how this is different. But my personal read of what Sam Altman has been saying about this is that it was sort of weaselly in its language in talking about that they believe that there is a need for human oversight of any potential lethal autonomous weapon systems. Well, the DoD already has a policy that there needs to be appropriate human oversight of any potential lethal autonomous weapons systems. But what that means in DoD policy is humans need to supervise the training of the AI. They need to test the AI. There's more red tape for a potential lethal autonomous weapon system than there would be for a different weapon system. And that ultimately then anything that that weapon system does, some particular operator in the system is ultimately accountable for that legally. But that's not what Anthropic was making their problem. They were saying, "Don't use these systems at all with our current models. We cannot consent to our models being integrated into these sorts of things." And so I very much read Altman's statement as not in keeping with the same red line, which, but it's interesting that he kind of made it sound as though it were. Like, I think he, I mean, these other labs are in a weird place because Anthropic is like getting a lot of plaudits out there right now for sticking to their principles on this and they don't want to seem like scabs or something like that, right? I mean, what's going on in the mind of your median OpenAI engineer right now as they're watching these things play out at the leadership level?
Totally. I think that they don't really know what to think right now. And again, I of course haven't spoken with a lot of people, just some, but it seems like, you know, they're just kind of listening to what Sam is telling them and then they're not really allowed to see the actual terms. So, they kind of don't know what to think. They're trying to gather information. You know, Anthropic employees, it seems the same thing. Like they're like, "Well, okay, did OpenAI just get the same deal we were fighting against? How did they do that?" And then they're thinking, "Okay, I guess not." Especially because of the wording, like you said. It seems like from the reports I've been seeing, Sam may have agreed to human responsibility for lethal autonomous weapons, meaning that, yeah, that could come after the fact, maybe, you know, not before. Anthropic was pushing for something, like you said, not at all right now, maybe later, and kind of figuring out the terms as the technology progressed. The other funny thing is I think Dario, the CEO of Anthropic, has been painted as like, you know, kind of just like an anti-war hero in this way. But like you mentioned, Claude has been used in the DoD for a long time and it was pretty much the most trusted technology of its kind. The other funny thing is that in his statement, Dario's statement a couple days ago, he mentioned that he is totally fine with lethal autonomous weapons, he just isn't fine with them right now. And he even offered to speed up the R&D on that technology with the Pentagon and said, "Hey, I'll work with you to get these systems up to par with where I feel comfortable saying you can use them." But they apparently, according to him, did not take him up on that. But yeah, it's been interesting how it's been painted kind of as a black and white thing when really, you know, he's a lot more okay with lethal autonomous weapons than the average person might think, just not right now.
Yeah. I mean, it's almost hard to communicate to people who have not been following this chapter and verse just how strange all of the different lines are because, like you say, I mean, this is like the 99% to where Hegseth wanted it to be, like defense contractor AI company up until now. Like they were hand-in-glove with the DoD in ways that maybe other labs would have liked to have been, but were nowhere near integrated with the Pentagon at the same level as Anthropic had been. And yet you get this. I mean, all the rhetoric coming out of the Defense Department right now is just 100%, you know, Dario Amodei has a god complex and he wants the United States to fail and he wants China to succeed. It's a radical left company. And so you're left in this weird place where, like, Anthropic, one, is not a radical company. They're not really a left company at all. I mean, this stuff doesn't map on perfectly to like just sort of US domestic political stuff, but they were happy to participate in, you know, national security stuff. And yet, you know, I guess they do really have these genuine red lines that they're willing to let the company suffer rather than violate. So, I guess like it's not like people are wrong to credit them with that, even though it's not true that they're like, you know, some sort of like they're not going around putting flowers in the barrels of the guns, right? I mean, that's not what this company's all about. Very, very weird, very, very strange situation. Let's talk just a little bit about what is coming next. I mean, Anthropic has said they're going to challenge this supply chain risk designation. I don't really understand, maybe this is not your beat either. Do you know what that would look like? What we're expected to see in terms of legal challenges going forward.
