Alix: [00:00:00] Hey there, I'm Alix Dunn, and welcome back to Computer Says Kill, a series exploring the people, politics, and systems that have ushered AI into the business of war. On today's installment, I'm joined by Maddy Batt, a legal fellow at TechJustice.law, who recently filed an amicus brief in the ongoing legal fight between Anthropic and the US government.

If you have at all been paying attention to AI and warfare, you have seen that Anthropic drew some really milquetoast red lines around how its AI systems could be used in military contexts, and the Trump administration totally freaked out and possibly retaliated, which is what the court case is about.

But in the middle of all of this, the technology that Anthropic has sold into the US military has been used in the illegal invasion in Venezuela and in the illegal invasion in Iran. Um, and in that illegal invasion in Iran, the US military struck a school and killed a hundred and seventy-five kids. And somehow, even [00:01:00] amidst all of that, um, Anthropic kinda came out looking like a good guy in all of this, even though they willingly partnered with the Department of War in a Trump administration that is committing war crimes like it's going out of style.

So today, we're gonna unpack all the ways that Anthropic is not the good guy, because behind this PR campaign, the company is deeply entangled in the business of war, just last week partnered with SpaceX to buy up all the compute in the Colossus I facility, which if you don't know, is currently poisoning people in Memphis.

And Anthropic saw no problem, um, taking up all the space in that data center and partnering with Musk, who is currently being sued by the NAACP for violating the Clean Air Act with the methane gas turbines that power that data center. So Anthropic is, you know, participating in environmental racism in Memphis, facilitating a humanitarian crisis halfway around the world, all the while cosplaying as this ethical consumer choice on their home turf.

So let's dig into it with Mattie Batt.[00:02:00]

Maddy: I'm Madeline Batt. I am a legal fellow at TechJustice.law. We are a legal nonprofit that fights for both consumer rights and human rights against harms caused by big tech.

Alix: Can we talk a little bit about what chatbots are doing in US military systems and a little bit of, like, foundation laying for people who may not be as familiar as you are about how these technologies are being Used in war.

So like, what was Anthropic contracted to provide the department in war? Yeah,

Maddy: absolutely. That's a great question, and in order to answer it, I actually need to take a little bit of a step back and talk about what AI- Love that ... more generally is doing on US military systems. And the answer to that question, at least for our purposes, is accelerating the kill chain to ever faster speeds.

And you already had an episode with Matt Mahmoudi where you talked about this, but for folks who may not have gotten to it yet, the kill chain is [00:03:00] a remarkably honest, I would say, military term for the process of going from identifying a particular person as a potential target, to assessing the impact and legality of a potential strike on them, to planning the strike, to ultimately killing that person.

And in the past, that could take days or weeks, maybe months for a high-profile strike. But as different aspects of the kill chain have been increasingly automated using AI, it can happen in a matter of seconds. And this process of accelerating the kill chain using AI predates the chatbot revolution. But once LLMs came onto the scene, they were integrated into that process in a way that, to caveat, it's hard from the outside to know exactly which aspects of an AI targeting system are attributable to LLMs versus other types of AI [00:04:00] algorithms.

But from the reports we do have, it seems that their role is both Substantive and significant. So for example, in Katrina Manson's really deeply reported book on Project Maven, which is the program that developed the AI targeting system that the US uses and that Claude has been embedded within, she cites military officials who say that the integration of LLMs into Maven increased the already AI-accelerated targeting process fivefold.

And she talks to the former head of Maven, Sean Battye, who talks about LLMs being able to play a role in identifying or prioritizing targets from among a list of thousands that have been identified by other forms of AI algorithms. Particularly troublingly from my perspective, he refers to the possibility of LLMs assessing the intentions of particular targets or recommending counter [00:05:00] actions even.

And so when Anthropic says in its court papers that Claude is playing a significant role in military operations, the reports that we have suggest that they're not exaggerating. And that raises a lot of concerns, both from the perspective of introducing error into targeting processes, as well as more generally just the capacity that AI generally and LLMs specifically have to scale up warfare in a really transformative and devastating way.

Alix: Most people, when they think about LLMs, think about, I feel like consumer-facing uses, so like, "Oh, I'm gonna make a workout routine," or like, I don't know, look up a recipe or something, or using it for almost like basic Google search. And so I feel like the relation to mistakes is really different in those contexts that I think most people think about versus the contexts you're talking about where, where mistakes obviously have [00:06:00] a more negative effect.

Do you wanna talk a little bit about what we know about are these technologies significantly better when they're deployed in military context? Do we know? I assume they're not, 'cause it's all on the same underlying type of technology, which is probabilistic and not deterministic, which means that they just will always make mistakes.

But, like, do you wanna talk a little bit about what we know about whether these systems work? Yeah,

Maddy: definitely. And I'll say we have a very good source, Anthropic itself, confirming for us that they are still making mistakes when they are used in this kind of system. So in its court case where it talks about its opposition to Claude being used in lethal autonomous weapons, it refers to the fact that, as you say, LLMs are probabilistic, and so they will inevitably and always make mistakes, and there's nothing different about Claude Gov that prevents that possibility.

And so I think there's real concern [00:07:00] that These mistakes will ultimately cost human lives as LLMs are deployed in targeting systems. I think it's also worth highlighting that these mistakes won't fall evenly across populations. LLMs have been documented to show racial, gender biases drawn from their training data that has swept in so much of the bias that exists on the internet.

And then the limited research that we have so far on how LLMs perform in war-like context is also really concerning. There was one study that included not Claude Gov, which is the model that Anthropic has provided to the US government, but a different Claude model that is publicly available, and it found that LLMs recommend nuclear strikes in 95% of simulated war games, which obviously should give us significant pause about deploying these [00:08:00] products in military systems.

Alix: Yeah. It's such a scary statistic when you put it in that context. If when you juxtapose the speed with which an AI system will recommend such a dramatic humanity ending potentially step with also the fact that it's making mistakes, so like there's kind of like a swirl there of both judgment in terms of decision-making and also accuracy, which I think is really...

Like it sounds like both of them are bad. Yeah,

Maddy: I agree. I also think there's potentially a deeper issue around just fitness for purpose as a preliminary matter. You know, again, in Manson's book, she talks to military officers who refer to LLMs and AI targeting systems as a reasoning layer. But in spite of the kind of sci-fi inspired marketing hype we get around LLMs, you know, they're not reasoning in the way that humans reason.

They're generating predictions [00:09:00] based on patterns they've drawn from training data. And so for an answer that an LLM provides to be useful in a given situation, the data that it's trained on has to be predictive of that answer. And for many of the kinds of questions that you might imagine a military asking a chatbot, it's hard to believe that there would be data that would produce an output that would be useful or reliable in that situation, just because of how dynamic war is and the degree to which a previous conflict won't necessarily be predictive of the way that people or opposing armies or civilians will operate in a new context.

Alix: It's also making me think about, I think it was Abeba Birhane that's the, that's the first person that ever explained probabilistic systems to me by saying you can ask a machine a question and you get an answer And you will never get that answer again. [00:10:00] You will never get that exact same answer, which I feel like in a war context, you would want to get the same answer, like the, in the same context, you would want multiple people to look at that context and say, "In this situation, this is the appropriate action."

You would want it to be able to answer consistently based on a set of understandings of the rules of war and, and all these other contextual factors. Okay, so everything's going really well. They buy these Anthropic technologies. They use them in the Venezuelan invasion. They use them in the Iran invasion.

We don't have all the details about how, but we know that Anthropic at least was behind the scenes being like, "Uh, maybe, uh, we don't want these technologies used in that way," or maybe, "Oh God, this is a public relations disaster that we've essentially partnered with an authoritarian state that's now like invading countries willy-nilly and buying our products to do it, and that doesn't look good."

So how did we end up in a situation where Anthropic is suing the US government? Yeah.

Maddy: So Anthropic wanted [00:11:00] language in its contract with the Department of War prohibiting the use of its technology for lethal autonomous weapons and mass domestic surveillance, and the department was refusing. Um, it didn't want any contractual constraints on how it could use AI, and it was insisting on only an all lawful uses kind of limitation.

When Anthropic very publicly put its foot down around this issue, the Pentagon designated them a supply chain risk, um, which was a highly unusual move for a domestic company, effectively blacklisting them from government work, and in a post online, Pete Hegseth was suggesting that they would also be blacklisted from work with other private contractors of the Department of War.

So Anthropic, I think persuasively argues that that was unlawful retaliation and sued to get the designation removed. [00:12:00]

Alix: So it was a, Anthropic said, "We don't want our technology to be used in mass domestic surveillance," and then in autonomous weaponry. So putting a model on a thing that's basically making a direct determination for use of lethal force or using Anthropic for mass...

W- why would the depart... I mean, this is an aside, but like why would the Department of War be interested in mass domestic surveillance?

Maddy: Um, that's a good question. The Department of War, in my view, is interested in anything that gives it additional power, and so I think the idea of any sort of external constraint on the way it could use its technologies would be anathema to the kind of domination-forward way that it operates.

Anthropic had previously had those written into its contract, and those were kind of red lines that the company had decided on as part of its government work. And then as [00:13:00] part of the Department of War's push for AI adoption that Pete Hegseth was really spearheading, they demanded that all contracts related to AI procurement include this all lawful use language, and that there be no external constraints on the way the technology could be used.

And so it could be that there were specific use cases that they had in mind. I definitely wouldn't put it past them, but it also is possible that it was related to this kind of contractual push to ensure that they had no constraints on the way that they were using their technology.

Alix: Interesting. Okay. So they sue for these reasons or these stated reasons, um, that they've been retaliated against because they've put together, like, very sensible red lines that any normal government agency would immediately say, "Of course we, we won't use your technology in that way," or, "We'll honor the contract," et cetera.

And then there's this retaliation for supply chain risk, then they sue. And then there's this [00:14:00] like, I don't know, 10-day set of stories and news cycles about everyone kind of coming to the defense of Anthropic and saying, "These are the heroes of our time," while we're also processing that in the Iran invasion in the first 36 hours, there was an airstrike, a US airstrike on a school where I think 175 kids, um, and teachers were killed by this airstrike.

So you have real victims of these airstrikes, and then somehow Anthropic becomes this kind of centerpiece of victimhood in this fight, which I found really disorienting and kind of gross, to be honest. But do you wanna talk a little bit about how that played out? Like why do you think we ended up in a situation where you had like progressive voices and even some people from civil society being like, "Poor Anthropic, this is so terrible.

We need to fix this"?

Maddy: I think it really comes down to the moment we're living in where tech companies have pretty uniformly been [00:15:00] gleefully kissing the ring of our aspiring authoritarian leader and reaping the rewards of that compliance. And so to see any tech company providing any sort of pushback against the federal government People wanted to celebrate it, and they wanted to defend against any possible negative repercussions.

And some of that, I think, comes from having a really low bar given the moment that we're in, and some of it comes from seeing Anthropic's pushback as a potential source of leverage against other companies that people are trying to push to resist the Trump administration as well. Um, but I completely agree with you with the kind of disorienting nature of treating Anthropic as a victim in a moment where its technology was being used to kill civilians abroad.

I think by losing sight of the substance of [00:16:00] the disagreement, we risk acquiescing to the use of AI in war in ways that none of us should be okay with, and that I think many of the people who came to Anthropic's defense wouldn't be okay with. I think it's also a really important moment to push back on this sort of foundational myth that we see in the development of generative AI that definitely OpenAI really pushed, at least in its sort of origins, and that Anthropic is continuing to really push, which is that it's a moral good to develop a powerful technology that you know can do lots of harm because if you develop it as quickly as possible, you can be in control of it, and then you can make sure it's used in an ethical way.

Um, and what we saw with the Anthropic Pentagon dispute is Anthropic trying to live by that central [00:17:00] animating philosophy of generative AI development in the US. And what happened was that, one, the constraints that they sought to put on the technology were totally inadequate, and two, when they tried to exercise the constraints, it ultimately didn't really matter to the way that the Pentagon has been able to use AI because as soon as they withdrew the tech, their competitor, OpenAI, which has been pushed forward in part by competition with Anthropic, filled that gap and provided the technology on the terms that Anthropic was refusing.

And so I think by celebrating Anthropic's pushback against the federal government, we also are losing out on an opportunity to look the falsehoods of this sort of foundational myth in the face.

Alix: I was like loosely following, maybe it was a week ago when, um, classified networks AI agreement. The War [00:18:00] Department announces agreements with leading AI companies to deploy capabilities on classified networks, and they partnered with SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle all at once And they made this press release, which makes me think they saw what happened where, like, one company got outsized influence and then defended in the public arena.

OpenAI then was vilified for kind of going right behind them and signing another deal, and rather than make it possible for there to be public picking off of individual companies, they were like, "We have to create a permission structure, so let's just, like, partner with everyone all at once so that we don't have to deal with singular companies having to deal with PR blowback."

But I also don't really understand it. Okay. So, um, how did you guys get involved in the case? Yeah,

Maddy: so we filed an amicus brief, um, because we were very concerned about the exact issue you were raising of the way that public debate had sort of shifted away from the actual [00:19:00] substance of the harm that was being caused already by AI and war toward defending Anthropic, and as a result, aligning with its mediocre red lines.

So we argued that even without full autonomy, Anthropic's tech is already facilitating, um, grave humanitarian crises in Iran and violations of the law. And I think what's important to understand here is the way that the adoption of AI in military systems has been driven by this speed wins ethos. That's language that comes from the AI memo that Pete Hegseth put forward.

The idea there is that if you are analyzing intelligence, making decisions, and ultimately killing faster than your enemy, then you'll be able to defeat them. And by turning to AI, every aspect of the kill chain [00:20:00] has been accelerated to extreme degrees, to the point that the human at the end of the chain who is making the final, kind of pulling the metaphorical trigger as it were, has become the slowest part of the chain.

And any time that a human being takes to assess a strike, to review the AI outputs that they've been provided requires compromising on the fundamental purpose that motivated adopting AI into the kill chain in- And so what we see is just rubber-stamping in practice. This is coming as well from reports based on the speed at which the US and Israel have been striking in Iran.

I believe it was over 1,750 strikes in the first 24 hours, for example, which is just an unprecedented speed to be approving strikes. And so [00:21:00] keeping in mind all of the issues related to error that we already discussed, as well as just the capacity for this technology to facilitate harm at immense speed and scale, the idea that humans are just readily rubber-stamping AI outputs is incredibly concerning.

It risks mass civilian harm, and it also means that the US is potentially precluded from making the kind of pre-strike assessments that it's required to under international law.

Alix: Yeah, you can't outsource that. Like, uh, nor should we be trying, and I find it so strange. I still can't picture what the end result is aside from scaled violence.

You guys are advocating that there should be the appropriate human agency at the appropriate time before lethal force is taken because there's a tremendous amount of international humanitarian law that spells out the [00:22:00] rules with a lot of that that seem to be ignored, and the absence of rule following usually means that a crime has been, has been committed.

Like, how do you think about Anthropic's partnership with the US Department of War and more generally AI companies partnering with the US government and committing these crimes? Like, how... Do you wanna talk a little bit about what we know in terms of is AI enabling more war crimes of the US administration?

Like, how, how do you think about that? Yeah.

Maddy: So AI is allowing the US to kill at unprecedented speed and scale. And from what we see, the US is striking at speeds that really seem to preclude the kind of pre-strike assessments that are required Under the Geneva Conventions. So for example, there are these cardinal principles of distinction which requires the US to distinguish between civilians and combatants and check to make [00:23:00] sure that it's actually killing combatants and not civilians.

There are the principles of necessity and proportionality, which require the US to assess that the strike that it's making is justified by military necessity and assess to ensure that it won't result in disproportionate civilian harm. And those standards afford the US discretion, but they do require assessment.

And when human operators are just rubber-stamping an output from an AI system that we know is unreliable because Anthropic told us so in its court papers, it's hard to see how that could comply with the laws of war. And that would also implicate domestic law as well because the US has codified serious violations of the Geneva Conventions under the War Crimes Act.

And so I see the role of AI here as really [00:24:00] significant for facilitating that level of speed that allows the military to move faster than the assessments of international law require. I also think it's worth noting that the US has been very open about its, I would say, disdain for international law and the Geneva Conventions.

Pete Hegseth has referred to the Geneva Conventions as stupid rules of engagement.

Alix: For like women. Yeah. They're-

Maddy: Yes. It's actually emasculating to abide by the Geneva Conventions. And so providing technologies that allow strikes to be executed in a matter of seconds when Hegseth has disdained the Geneva Conventions, when Trump has posted online about wiping out the entirety of Iran's civilization, I think there's a lot of question there about why you would choose [00:25:00] to facilitate violence at scale by a military that has been so open about not abiding by international law.

Alix: Yeah. I feel like there's two parts of this. The first is, is integrating AI increasing the likelihood of war crimes or enabling war crimes. And then there's this other piece to it, which is, is the introduction of AI actually making it harder to tell what's happening in a way that decreases the chances that accountability down the line will happen?

Because basically, who made the decision? Was it the model or the company or a senior military official or a someone that's more junior or some combination therein? And I'm wondering, 'cause the other way of thinking about this is that they're essentially creating much more direct digital traces of how decisions are made, because I imagine a lot of those decisions got made and weren't documented when a person thought about it in their head.

But now we have... Like, to what extent could you imagine a scenario where digitizing the decision-making [00:26:00] infrastructure of a military actually increases the chances of there being accountability later versus it just being this like kicking up dust so it's really hard to distinguish between the kind of fog, basically creating the fog of war using AI as a cover?

Maddy: Yeah. I think it absolutely creates more concerns for accountability than possibilities for accountability. You know, AI is famously a black box. And so while you're getting outputs, you're not necessarily getting an explanation of how that output was created. That's true even though there are reports that Maven Smart System is automatically generating legal justifications for strikes that it recommends, which continues to support the concern that these assessments are being outsourced to technologies that aren't really suited to make them.

But I think it's worth talking about accountability issues for lethal autonomous weapons on the one hand, and then the sort of semi-autonomous systems that are often referred to as [00:27:00] AI decision support systems. Because on the lethal autonomous weapons side, which is the example of, you know, say a drone that can go out into the world, identify a person as a target, decide to kill that person, and then kill that person without any human input, legally there is a significant accountability gap there because there's a real possibility that there is no human who possesses the required scienter or kind of legally culpable state of mind to be held accountable for that death if the autonomous weapon kills a civilian.

For systems like Maven Smart System, where a human is still making the final call, even if they are rubber-stamping what an AI has provided to them, I would not cede that there's kind of a legal accountability gap there. I would say that the person who kind of accepts the AI recommendation should still be legally responsible for that [00:28:00] decision.

But I think there's a real kind of moral accountability gap that we end up with in these situations because obviously It's imperfect. We still have atrocities that occur, but humans, one hopes, feel a moral responsibility when they decide to take a human life. And I think that does provide some sort of meaningful check on some of our worst impulses in war.

And when a human is able to abdicate their sense of moral responsibility when deciding to take a life because their only responsibility is to click accept on an AI recommendation, I feel a lot of concern for the moral space that we've landed in in terms of a military being able to effectuate death without anyone feeling [00:29:00] that it weighs on their conscience.

Alix: There is a moral philosopher, Linda Eggert. She has written about autonomous weaponry should never be allowed because it can't experience regret, is like the way that she frames it, and I have always found that... I like, I go back to that as like a, maybe not the perfect check or like legal threshold, but just as a moral reasoning to me that really resonates that like I don't want something that can't feel shame long term, um, to be making those kinds of consequential decisions.

I also thought, I mean, what's the Stop Killer Robots campaign doing? I thought that like, I feel like DeepMind early on, at least among the labs, made some pretty, I thought, clear statements that they never wanted the models they were building to be deployed in autonomous lethal weaponry. But is it now? Is that happening?

So

Maddy: this is, I think, where there are some challenges around the distinction between lethal autonomous [00:30:00] weaponry and what we call semi-autonomous military AI. So Maven Smart System would generally not be considered lethal autonomous weaponry. It would usually be referred to as an AI decision support system.

And on paper, there is a clear distinction between those two types of technologies because a lethal autonomous weapon is acting totally without human input, whereas an AI decision support system, a human is still in control, is making the decision to act and acting, but is just receiving intelligence analysis support and suggestions from AI that it can take or leave.

But what we see in practice is that AI can still end up driving the decision-making process because of that speed wins impulse, where taking time for further human [00:31:00] consideration of a strike is compromising on the ability to move ever faster that AI adoption is intended to further. And of course, in many cases, it's not actually possible for a human to meaningfully review the AI recommendation because it's based on terabytes of data that a human could never kind of review and comprehend in order to, um, meaningfully assess whether the AI output is accurate.

And so Anthropic and potentially other companies are staking out a claim specifically related to how their technology will be used in lethal autonomous weapons while embracing this other form of using AI in war that on paper is different, but in practice still uses AI to drive machine decision-making over who lives and who dies.

Alix: I wanna return to this question [00:32:00] of who the good guys are, um, because I think I'm currently feeling even clearer than I was when we had a news cycle that juxtaposed the death of these kids in this school and then Anthropic, and then Anthropic becomes the victim. But very recently, Anthropic signed a partnership deal with Elon Musk and SpaceX to take over all of the compute capacity of Colossus-1 in Memphis, which is basically everyone knows is like the most disgusting neighborhood poisoning data center.

And everyone, I think, has kind of accepted that Musk is this like environmentally racist, just like disaster, but then seeing Anthropic happily plug all of their systems into the exact same compute capacity that is now poisoning those kids, to me it's much for muchness. Like, it's basically just like Anthropic...

There is no difference in my mind between Anthropic and [00:33:00] OpenAI, and a- all these companies now in my head are just like It's the same. But what do we do? 'Cause, uh, 'cause if basically if you just lump them all together, they have so many resources, and, like, I can see the attraction of trying to, like, pick them off one by one to be like, "Anthropic is better than you," in a way that, like, in some ways makes Sam Altman look bad in a way that is maybe helpful in the public consciousness to try and push on that.

Or I can see some strategic advantage of differentiating between the companies because then you have the chance of maybe influencing their behavior in a kind of game theory sense. But ultimately, it's, like, a dog shit sector that's, like, doing anything possible to, like, make it through this gulf of tech transfer where they have no business model yet, but they have kind of an interesting technology that maybe in some cases does some interesting things that require a tremendous amount of resources to make happen.

Um, like, how do you, how do you think about that? Like, how do you think about, obviously you filed this amicus brief, obviously you're interested in helping set some legal precedent around the [00:34:00] specificity of the claims that these companies can make and, like, how they're positioned as technology providers for the US government.

Like, I, I, I see you working in the detail of administrative controls, but, like, how do you feel more broadly about who the good guys are and, like, how we use that to our advantage to improve outcomes? Yeah.

Maddy: So I think the fact that you talked about data centers is really interesting, because I feel like that is actually a really nice way into the way that I think about this, which is that the fact that these companies are so powerful and so far-reaching into different aspects of our lives also means that there are a lot of different places to resist them.

And so for me, when I think about the good guys, I'm thinking about all of the different pockets of struggle and resistance that are pushing back against the exploitation of tech oligarchy in kind of different places within their supply chain. And [00:35:00] data centers are a great example of that, where there have been incredibly inspiring victories, um, of people at the local level pushing back against exploitation in their communities that down the line is fueling AI militarism as much as it's fueling all of the labor problems we see associated with AI, as much as it's fueling the consumer protection violations that we're seeing with AI products as well.

And so for me, I don't think it's useful to distinguish between AI companies and say, you know, Anthropic insisted on different contractual terms in its defense contract with Trump's Department of War, and therefore it's a good guy as compared to OpenAI. Because I think we just need to set the bar higher than any of these AI companies have been willing to reach for.

But we can be clear-eyed and [00:36:00] optimistic about the interconnectedness of different struggles, and look for inspiration from places where resistance is happening that might seem distinct from the issues that we're working on ourselves, but actually are all part of the kind of mammoth effort that it will take from all of us to push on, on different parts of these companies' supply chains, different parts of the web that is these tech companies' power in order to take back control from them and the sort of exploitative model that they've developed.

Alix: I like that a lot, and I think looking at it as a system is really important, 'cause I think it's when you try and, like, break off individual instances or in- of companies, it's just not nearly as effective at shifting the overall system, which is, I think, what needs to happen. But, like, law usually does zero in on, like, particular entities that can be held liable or, like, [00:37:00] particular lines that can be articulated with some level of precision.

How do you see the legal profession playing a role in some of this system change work? Yeah.

Maddy: So I think lawyers have a valuable role to play in foregrounding and enforcing the norms that we already have that constrain tech companies. There's this narrative that I think is honestly quite useful to them that emerging technology is operating in this legal vacuum because tech is moving faster than the law.

And I won't say that there are no areas where emerging technology doesn't create problems for legal accountability or that it wouldn't be useful to have kind of clearer legal norms around AI and emerging tech, because both of those things are true. But it's also true that if you commit a war crime with AI, you've still committed a war crime.

We need to resist the idea of a complete legal vacuum existing around emerging [00:38:00] technologies, um, because it ultimately serves tech companies who want to be the ones to independently set the standards for how their tech is used. And as we've seen with this dispute with Anthropic, even the, quote-unquote, "good guys," though I think we've rightly been problematizing that today, are not setting standards that are even in line with our existing humanitarian laws.

And also from a narrative perspective of helping reclaim from tech companies this idea of a legal vacuum and say, "No, even though you're using novel technology, you're still accountable to the laws we already have and the human rights norms that we already have." And so I think lawyers have an important role to play in that kind of narrative fight as well.

Alix: That makes loads of sense to me. And I will say one of the most moving moments I've had in the last, like, I don't know, maybe two years, was [00:39:00] watching the case get presented by the South African delegation against Israel. Um, and watching them, like, systematically in, like, a no spin, no media narrative environment in, like, a legal-- in a courtroom where it's, like, so much harder to bend things and kind of step outside of a, an evidentiary framework.

Like, watching that case get so methodically laid out was just so satisfying because there's, like, nowhere, nowhere they could run. Um, and it was-- I just found it really-- It makes me think that even though it probably will take a very long time, that it probably will ultimately not be as satisfying as we would want, there are going to be these moments where they can't get out of this forever.

Like, there will be moments of accountability that are coming, and I am excited for that to happen.

Maddy: I agree, and I'm biased within the legal profession, but I do think that it means something to go into a courtroom and say that [00:40:00] you have violated norms and obligations that you hold to me. I think that has power.

Alix: Okay. I think that's a great place to leave it. Thank you so much. This was wonderful. And, uh, thank you also for taking the time to file the amicus brief because it was a good read.

Maddy: Thank you so much. Thank you for having me.

Alix: Thank you to Maddie for providing a legal perspective on accountability at a moment when accountability feels more important than ever and more absent than ever. Next week, I'm gonna be joined by Anthony Loewenstein, author of "The Palestine Laboratory," fantastic book. If you have not yet read it, please go buy it immediately.

Um, he's gonna break down exactly how Israel has used Palestine as a testing ground for weapons of war and the history of Israel's partnership with the military industrial complex. So stay tuned, and I'll see you next week. And as usual, I wanna thank Georgia Iacovou, Kushal Dev, Marion Wellington, Sarah Myles, Van Newman, and Zoe Trout for working to produce this episode and [00:41:00] series.

‍