Jack Murphy (2:30)

Oh, shit. I mean, it's a really bad scenario to have Americans, especially one or two lone Americans, behind enemy lines. That's like the worst case scenario if you're a Soldier escaping and evading a much larger force. And whether or not you get picked up or you make it, the friendly lines is debatable. So I mean, yeah, it's a scary, tense situation. I mean, mostly for the two gentlemen involved, of course, but I mean, I think the rest of the country was holding their breath too. Aviation is not necessarily my speciality, but as I understand it, the ejection seats for the weapon system officer and the pilot go in opposite directions. So they were actually ejected in opposite directions, as I understand it. And that accounts for them landing somewhat separately from one another. I mean, a lot of this information, more detail is going to come out in subsequent days and weeks.

Venetia Rainey (3:24)

And what's going on behind the scenes in CENTCOM after they hear this. Just, just walk us through the parts of the US military that are kicking into action. You, the elite combat search and rescue team and the presumably very well rehearsed plans that are being.

Jack Murphy (3:36)

Yes, there are contingency plans for recovery personnel. And what immediately happens is they launch combat search and rescue. They are looking for these gentlemen for these, the pilot and the, the wso, the wizo they're looking for. That's the weapons weapon system operator. Yeah, he's the backseater in the F15. So they launch combat search and rescue. And this is actually an interesting detail. I don't think I've talked about this yet. The Cesar guys went out on helicopters and they punched out so quickly that we didn't get the electronic warfare countermeasures up that we normally have for our air campaign. So these guys were just so gung ho to go in there and get this guy. They went and their support aircraft were struggling to catch up with them because they were so far out ahead.

Venetia Rainey (4:26)

So this is the first time in the war that a plane has been downed. It's quite a rare occurrence in general. Right. For American fighter jets. A huge potential propaganda win for Iran if they find the pilots. And a moment of peril for America, as you say. Just what do you think shot the plane down? We heard at the time that Iran's top joint military command said it was a new Iranian air defense system. What did you make of that?

Jack Murphy (4:46)

Yeah, I don't know if it was a new Iranian defense system. I mean, what they were talking about, using infrared to target aircraft is super old. It's been around for a long time. There's some speculation about what is shooting down our drones and now aircraft. There's a Iranian built missile, I think it's called the Thayer, but it's a Surface to air missile. And then there's also been Russian S300 and S400 surface to air systems. How many of those are still operational? I don't know. So I think it's going to remain to be seen exactly which system took it out.

Venetia Rainey (5:22)

It's a bit of a problem. Right, because we've had Trump repeatedly saying, and Hegseth be hegseth Secretary of Defense saying that, you know, the missile systems have all been taken out.

Heidi Klaff (5:30)

Yeah.

Venetia Rainey (5:30)

Clearly, either they had a missile system or something that could take down an American fighter jet, which is quite.

Jack Murphy (5:36)

Well, yeah, they took it. They took down the fighter jet. Two Reapers Reaper drones were taken down during the course of this rescue mission, and our helicopters got the hell shot out of them. So this is contested airspace. Like, we don't have that kind of air superiority that America has been used to dealing with through the war on terror in Afghanistan and Iraq.

Venetia Rainey (5:55)

Okay, so talk us through that first rescue mission. At around 4pm GMT, we start hearing reports that the pilot has been rescued. But it's clear that it was quite a messy operation, as you've alluded to. So just talk.

Jack Murphy (6:06)

I mean, I don't know that I have the minute by minute on it, but the aircraft had to make some midair refueling runs in broad daylight, flying very slowly over Iranian airspace. And I'm sure that was hair raising for everyone involved. And yeah, my understanding is that they took a substantial amount of ground fire from the police and the other search parties, Iranian regime authorities that were out there. But it does seem like they picked up the pilot relatively quickly.

Venetia Rainey (6:36)

And this was an A10 Warthog. Right. And two helicopters, I believe these were

Jack Murphy (6:40)

Pavelo Black Hawk helicopters that picked them up. The A10 was there to support it was part of their security element and that got shot up so badly. And an A10 is a very robust aircraft. I mean, it's. It's pretty rare for one of them to get shot out of the sky. The pilot had it under control but deemed that the aircraft was not landable. So he had to do a controlled ditch, essentially a controlled landing or a controlled crash maybe is the more accurate term. And then he got picked up and pulled out.

Venetia Rainey (7:11)

Two he ejects over the Gulf right outside of Iranian territory, I believe in Kuwait. Okay, so the other airman is still there, a colonel, we now know, weapons system officer, this guy is stuck somewhere in Iran. And we know Iran has launched search parties. There are videos on social media of hundreds of people heading up to this very mountainous area just to locate People, we're in southwestern Iran. It's very mountainous. I heard on the podcast that you're on someone describing it as like moon dust, like the is hard terrain. Iran's military is offered a bounty of around £50,000 to anyone who captures the crew member alive. What would this colonel be on the ground doing? What's the US Air Force's survival, evasion, resistance and escape program like?

Jack Murphy (7:51)

Oh, okay. So I mean, there's a lot that goes into that, but I think kind of some of the more clear and obvious things is you're escaping, innovating, so you're trying to remain concealed and not be caught. The other thing you're trying to do is to use your beacon or survival radio or whatever communications device. You have to be located by friendly forces as quickly as possible. The longer you spend out there, the worse your situation is going to get. You want to get picked up as soon as humanly possible. So I suspect that he was evading, that he was on the run in order to prevent his capture, while at the same time he was probably messing with his survival radio and whatever other signal devices he had. And I'll get a little bit more into that in a bit to try to get rescued now. I was told that he authenticated over the system he had quite late actually, just a few hours before the rescue mission was launched.

Venetia Rainey (8:45)

What does that mean?

Jack Murphy (8:46)

Authenticate some sort of secret authentication code that he would give over the radio to say, this is really me, I need help. And I don't know why that came in late. That's one of those things that will come out down the line, I guess, and it doesn't really matter in this context. I mean, the guy was under duress and they got to him as quickly as they could.

Venetia Rainey (9:07)

How important? I mean, just, I just want to sort of go behind the scenes again and you know, CENTCOM headquarters. How important would it have been for them to get this second guy out? The US's military's mantra is no man left behind, right?

Jack Murphy (9:18)

This would have been their number one priority. All assets would get shift to this. This would become the main focus of everybody at centcom. And you know, because this became a rescue mission, it got passed out to Joint Special Operations Command and the Joint Personnel Recovery center also involved in this. But it became like a no kidding Special Ops task force being sent in to repatriate this guy.

Venetia Rainey (9:42)

And I've seen some reporting that at first there was no sign that he was alive. The military described him as status unknown. How concerned were you at this point, we're sort of, you know, Saturday midnight.

Jack Murphy (9:51)

I mean, I'm very concerned. I think everyone was very concerned. However, by, you know, yesterday morning, the fact that he was not being put all over Iranian social media led me to believe that he was in one or two states. He was either still on the run or he had been taken in by friendly anti regime resistance elements and was being like hidden inside a barn or something like that, like they were taking care of him. I figured it had to be one of those two things, but I mean, obviously it's. It's still scary as hell.

Venetia Rainey (10:22)

Why didn't you think he might be dead?

Jack Murphy (10:27)

That's a possibility. I mean, you don't want to make that assumption. It is something that, you know, I thought about, but it seemed likely to me that he had a successful ejection from the aircraft. So he was most likely alive when he came down onto the ground. That was my suspicion.

Venetia Rainey (10:44)

So we now know that he hiked up a 7,000 foot ridgeline and wedged himself into a crevice where he hoped he would be safe until American forces found him. That's according to U.S. military officials. What was the search operation like to try and find?

Jack Murphy (10:56)

Yeah, my understanding is that the Iranians pushed a tremendous amount of assets into this region to look for this wiz up. I know for sure. I believe they're called the Basij. They're like the Iranian secret police or political police were out there looking for him. And that he apparently, from what we know, scaled up some cliff or mountain actually makes a lot of sense. It makes sense that he had evaded Capture for over 24 hours by using the terrain to his advantage. You know, most of Iran is flat and open, and he wouldn't have been too difficult to locate in terrain like that. But by going up into the mountains, even if they knew where he was, it would have become a lot harder for the Iranians to catch him and actually detain him. So that was a very smart move on his part to use the terrain to his advantage to buy time.

Venetia Rainey (11:47)

And the Iranians were searching on foot. Right. They don't have really any kind of air force.

Adam Wishart (11:51)

No.

Venetia Rainey (11:51)

Although I guess they could send out drones.

Jack Murphy (11:53)

Yeah. And I mean, of course they were trying to get into the area with vehicles, and we bombed some of the roads in order to prevent that, and we bombed some of the communications towers to stop them from talking to each other. So there were preparatory fires around this area.

Venetia Rainey (12:05)

I also read that there were Iranians on the ground who were hostile to the Regime, should we say? And that blockaded some of the roads. There was video.

Jack Murphy (12:13)

Yeah, yeah. A colleague of mine, John Hackett, was talking about that this area of Iran is pretty anti regime, and he was pointing out that they were deliberately causing a traffic jam in order to prevent the Iranian authorities from getting into that area. So, I mean, I'm sure everyone greatly appreciates their contribution. You know, if that happened. I hope it did.

Venetia Rainey (12:36)

So talk us through the actual operation to rescue the second M. And there's sort of two parts to it, right? Just talk us through how exactly they found him and physically got hold of him, and then the second part, how they actually got him out of Iran. We'll get to that in a second.

Jack Murphy (12:48)

So as far as I know, he got up on whatever communication, survival communication system he had. I know there's stories about the CIA going around and how the CIA did this big heroic thing, but I've had sources tell me that that's basically just political BS that they're, you know, spreading to journalists, trying to take credit, specifically Director Radcliffe trying to take credit for a military operation and make himself look good in front of his boss. So, I mean, maybe the CIA had some, you know, role, ancillary role in this. I don't want to discount that. They may have resistance networks on the ground. I don't want to discount that. But I don't think the Agency had quite as much to do with it as the Director might want the press to think. But anyway, he did come up on the radio and authenticate at a certain point, and that launched the JSOC task force to go in and get him. And at this point, it becomes a complex operation. This is a contingency operation. So just by the very nature of it, it's completely reactive. It's not preemptive. You have some plans in place, but you're having to react to the situation as it unfolds in front of you. And in retrospect, I think everyone involved did a terrific job.

Adam Wishart (13:58)

So

Jack Murphy (14:00)

the aircraft that they flew in was the C130 platform, which is one of the. It's the main hall cargo aircraft of the US military, essentially. I mean, the C17 is another big one, but the C130 is our kind of the mainstay. They flew in on that. There was four little bird helicopters in the task force package. Those would have been loaded into the back of One of the C130s, unloaded at a remote improvised desert airstrip, which the military calls a farp. It's a forward arming and refueling point. So it's essentially you're setting up like a little base, like a little ad hoc base somewhere where you can refuel aircraft so that they can get a little bit further out to do their mission and then come back and get more fuel or fly off or whatever the case may be. So they set up a FARC there, they get the little birds operational and ready to go, gassed up, and they launch to go and rescue the Wizzo. And this is an interesting detail that actually I just learned. So the Whittlebird aircraft are very small but nimble aircraft that can carry special operators on the side of them. They're kind of hanging on the outside, so to speak, sitting on, like, a bench. And there were Air Force special tactics guys and Seal Team 6 operators who were loaded onto these aircraft and flying out there to rescue this guy. As they got closer, about 9 of these bossage political police guys who are close to the wizzo were killed. Now, I don't know for sure if they were killed by the operators on the ground or if they were killed via close air support. I think that's another thing that will come out in subsequent days. But an interesting detail I heard was that the wizzo used his boxer shorts to signal to the helicopter pilots and the Navy seals, like, hey, hey, here I am, here I am. It's from a football team, an American football team. Has the logos on his boxers, and those boxers have been put on Iranian social media. The Iranians are showing them off. So, yeah, he used his boxers to great effect and then discarded them. They weren't coming with him on the ride home, I guess. So that part's a bit funny. But as I understand it, the little birds did take substantial gunfire during this operation. But everyone alive, they fly back to the farp. And this is where the mission threatens to come undone. As I understand it from a couple sources, there were mechanical problems with at least one of the C130 aircraft. What I was told was that there was a mechanical issue, that the aircraft had sunk into the ground, like, sunk into the earth, and the crew was trying to dig it out. Like, and I don't mean the entire airplane. I mean just the wheels. And it gets. The wheel sinks in enough to the wheel well, and now it can't move anywhere. So they're trying to dig it out. But we're getting to the point where the sun is coming up and these guys are deep behind enemy lines, so we need to get them out of there somehow, some way. And they're not it seems that they're not having much success digging out the C130. So the alternate plan that comes up is they fly what's called a Casa C295W. It's like a lighter, civilianized sort of version of the C130. It's a smaller, lighter aircraft. And they flew that into that airstrip to rescue these guys with what's called the Quick Reaction Force, which, again, it's a contingency. You have additional soldiers standing by ready to respond in case things get ugly out there. The QRF was coming from Delta Force, as I understand it. So you had both Seal Team 6 and Delta on this target. They get out there, and the decision is made. You know, they can only load two of the little birds onto this new aircraft, the C295. So they have to destroy two of the aircraft. Two of the helicopters have to be destroyed there at the farc. The two others are loaded into the airplane with passengers, and. And now there's two C130s on the ground that also had to be destroyed. Now, were they destroyed with, you know, the guys on the ground putting plastic explosives on them and demoing them, or. I'm also told that the Air Force dropped literally tons of munitions on these aircraft to destroy them after everyone had successfully taken off. Maybe both of those things.

Venetia Rainey (18:27)

And why is that so essential to destroy this aircraft? Just talk us through.

Jack Murphy (18:31)

Yeah, there's a couple reasons why you would want to do that. First and foremost is to deny the enemy that intelligence information. There's everything on those airplanes from, like, encrypted radios that have our ciphers and things like that on it, to classified navigation systems, classified air defense systems that can defeat or prohibit missile launches. And some of these things are very much classified systems. And if the enemy was able to capture them, they could begin to reverse engineer them and figure out how to defeat them so they would no longer be useful in combat anymore. And they would also, you know, if they're captured by the Iranians, you can totally believe that they will sell that technology or that information to the Russians and the Chinese as well. So that's the biggest reason. The other reason is to deny the enemy the use of, you know, of those aircraft for propaganda value. You know, one of the Blackhawks that went down in Somalia, the Somalians actually grew cactuses around it. And it's there to this day. It's like a miniature. I want to say it's a tourist attraction because there aren't too many tourists, but it's still there to this day. So, yeah, it's called blown in place bip. No one wants to do that, but it was deemed necessary at this point.

Venetia Rainey (19:45)

So they finally get out of Iranian airspace. Trump gets to put out this tweet saying, we got him. You broke the story before everyone else knew about it. You broke the story that the Wizzo had been rescued and that he was alive. You also said that there was a massive firefight on TGT and gas market. Iranians were actively looking for him in the area. I just want to put to you some alternative reporting that I've seen in the New York Times. They say that commandos did not engage in a firefight with enemy forces. And they cite US Officials describing the territory. And I know you mentioned this from Jonathan Hackett describing the territory where the airman was hiding as strongly opposed to the Iranian regime. And it's unclear how close Iranian forces ever got to the site. What do you make of that?

Jack Murphy (20:22)

Well, the Iranians got fairly close from what I'm told now. But again, were those guys engaged by commandos on the ground, or were they engaged by, like, a 10s or Apache helicopter? I actually don't know. But, I mean, there's. There's videos out there. I don't suggest looking for them, but there's videos of dead bodies.

Venetia Rainey (20:41)

So what, what are the unanswered questions for you that you'd still like to know about this operation? Details get in the coming days or weeks.

Jack Murphy (20:49)

I mean, this whole operation was incredibly complex, and I think a lot of lessons learned are going to come out of it. I think, you know, the bravery of a lot of these guys is going to be highlighted rightly so down the line, you know, I was told, and I think Trump, you know, said this himself in the social media post he made, but internally in the military, they were referring to this as an Easter miracle. And I mean, maybe they're not wrong. I mean, they kind of pulled off the impossible. To go behind enemy lines and snatch this guy out without losing anyone else is, you know, pretty incredible. And the technical and tactical proficiency of the US Military has been demonstrated over and over again. They're very good at what they do. They really are. The problem is that our policymakers are not as technically and tactically proficient as the military. They don't have the strategic foresight to really employ them in a way that advances American interests. And that's maybe the incorrect lesson from their success that sometimes our policymakers take away is that they see airstrikes and special ops as this sort of like, easy button that they can use whenever their foreign policy is failing. And as this operation shows, you have to be careful about how you employ American military power. It has to be used intelligently.

Venetia Rainey (22:13)

As you said, this is one of the most complex operations in American military history. Probably. What precedents are there for an operation like this?

Jack Murphy (22:21)

Well, the one that immediately comes to mind as being the closest parallel is Operation Eagle Claw, which was a 1980 failed effort by Delta Force to rescue American hostages being held at the American Embassy in Tehran. And that operation was also very complex. You had airplanes and helicopters launching from two separate locations and meeting at a remote improvised desert airstrip known as Desert One. The mission was to link up there. The helicopters would then fly the Delta operators to the outskirts of Tehran where they would then board vehicles that had already been procured and drive to the US Embassy and conduct a hostage rescue mission. The embassy grounds themselves were pretty huge and required, you know, pretty much every operator they had at the time. Unfortunately, they didn't make it that far. A helicopter collided with one of the fixed wing aircraft at Desert One and that went up in a fireball and killed some people. But actually, even before that happened, when the helicopters arrived, they had flown through a sandstorm and a bunch of them were no longer flyable. They weren't operational. And so the mission really had to be aborted at that point.

Venetia Rainey (23:32)

That was Jack Murphy, former U.S. special Forces and military journalist. Coming up after the break, what role is AI playing in the Iran war and beyond? Welcome back. You're listening to a special bank holiday episode of Iran, the latest. Now, shortly before the Iran war began, there was another huge story that we just didn't quite get around to covering. It was a massive row that erupted between AI Giant Anthropic and the Pentagon. Now this round was over the US Military wanting to use Anthropic's AI system, which is called Claude, and you might be familiar with it. They wanted to use this in autonomous weapons and mass surveillance. Anthropic said its systems weren't good enough yet and they refused to sign a new expanded contract. So Secretary of Defense Pete Hegseth responded by labeling Anthropic a supply chain risk. And he ordered all government agencies to immediately stop using the company's tools. The legal battle over that decision is still ongoing. But this raises some really important questions. What role is AI already playing in current wars from Iran to Ukraine? What impact is that having on the ground? And if we all got a say, what role should AI play in warfare? This is a live issue. Just this week, Israeli newspaper Haaretz revealed for the first time that Israel is using multiple AI systems for offensive purposes as well as defensive. And that includes suggesting targets, finding them and tracking them in real time. Now, AI's proponents argue that this makes war more efficient, more precise and quicker, and that it gives the West a key battlefield advantage in a time of rising conflicts and threats. I wanted to know if that's true, so I've been speaking to two experts in the field. Adam Wishart is a filmmaker who's just put out a documentary on Channel 4 called Click to Kill the AI War Machine, looking at how AI is being used by Ukraine, Israel and the US army in Europe. He's spoken to American commanders, Israeli whistleblowers and Ukrainian drone operators and will be sharing their insights into this contested area. My other guest is Heidi Klaff. She's chief AI Scientist at the AI Now Institute where she focuses on safety within defense and national security uses of such technology. She's previously worked at one of anthropic's main rivals, OpenAI, so she's a proper industry insider. Here's our conversation. Heidi and Adam, welcome to Iran. The latest. There's obviously a lot to get into about how AI is changing warfare, but I want to start in Iran. Heidi, can you explain exactly how we think AI is being used in that conflict?

Heidi Klaff (26:31)

Well, we're in Iran. We're actually seeing multiple types of AI being used. From the US side. We're seeing Palantir's Maven systems being used, which is powered by Anthropic's AI model Quad, that was used for recommending targets and contributed to 1000 strikes in the first 24 hours and 5000 targets within the first 10 days. And essentially the type of AI that we're talking about here is what we call a decision support system, which are essentially tools that bring together a lot of data, like satellite images, social media feeds, intercepted communication, and uses them to make military recommendations, including targeting recommendations. It's a tool that aggregates information and operates on every level of the kill chain. Now, this is very different from the type of AI used by Iran, for example, for shahed drones. And we've seen this type of AI also used in Ukrainian drones, which is a different conflict. But the reason why I bring that up is that those types of drones are actually very different from these decision support systems. They're what we call purpose built military models that are trained for specific tasks, that are much smaller in scale and are trained on far less data because of their task specific nature. And also they have a much more vetted supply chain. This is very different from the large language models like CLAUDE or Generative AI being used by the US which are these general purpose models with a commercial supply chain that can be compromised that are then repurposed for military applications.

Venetia Rainey (28:00)

So there's a few things that I want our audience to hold in their heads. They're decision support systems. That's really how the military uses AI to help it find targets. If we want to be super simplistic. Right. And the US army uses, as you mentioned there, a system called Maven, which is run by Palantir, which uses as the backend anthropics CLAUDE AI model. And our listeners can go away and type something into claude. You can have that app on your phone and ask it a question. What the US military is using is obviously something much bigger, more complicated. And as you said, it pulls together lots of different sources of information to provide much more complex advice on potential targets. Adam, how does this differ from how the kill chain used to work? Just give us a sense of how AI has crunched that up in a really short period of time.

Adam Wishart (28:44)

In my documentary, I interviewed Louis Moseley, who provided Palantir support for Ukraine, and also General Chris Donoghue, who's the land commander for NATO and the US troops in Europe. And they and their teams basically say that in the old days, human analysts would have to look at signals, satellite intelligence and figure out, for example, where a tank is. And so a human would have to go through all the dozens, perhaps or hundreds of signals and satellite images and figure out what is a tank and what is not a tank. And nowadays what is happening. As someone in my film describes, they fed loads of images of tanks into their large language model. And their large language model says that it can now identify what is a tank and what is not a tank. And so now they can feed in all the millions of satellite images, millions of mobile TE signals, anything that's scraped off the public and private webs. They can feed that all in these many different sorts of data. And they can say what looks like a tank. Or, for example, in my film, someone says that satellite imagery is very good for working out where something is, but it doesn't say what its importance is. And so if you overlay the satellite imagery with signals intelligence, you can work out not only what it is, but whether or not, for example, it's a Russian command center. And so all of this is now being done at a fast rate in Iran by the US military because the Maven smart system underpins everything. And it doesn't only underpin the target decision making process, it also contributes to where they should deploy resources, what munitions they should use, even quite banal stuff about whether this F16 needs a repair or not.

Venetia Rainey (30:24)

Heidi, you mentioned 1,000 targets were hit in the first 24 hours of the Iran conflict. Would that have been possible without AI? Would humans have been able to generate that many targets by going through all the information available?

Heidi Klaff (30:37)

No, but I think something that's missed out on is the lack of accuracy of these models. As Adam was saying, AI is very good at aggregating large amounts of information and providing inference on that data or giving pattern matching insight. But the problem here is that it's not very accurate. It can miss things, or it can label something as a military target that isn't one. So a good use case of how to use AI is giving it a huge amount of unstructured data, like Adam was saying, satellite images, and then exploring the recommendations they make. So it can be helpful in whittling down millions of data points into more relevant sets that a human operator can then sort through. But what's usually happening is that militaries try to take these recommendations and apply them to targeting directly for the sake of speed and decision making. But if you look at investigation on these systems when they're used for targeting, we're looking at 25 to 50% accuracy. And two years ago there was an investigation, investigation showing that MAVEN can oftentimes have 30% or less accuracy, which isn't really far from indiscriminate targeting. And I think that shows that although it can help you with understanding and sorting through your data better, it's not a substitute for human decision making given how inaccurate they have been for targeting. And I also think it's very dangerous that speed is somehow being sold to us as being strategic here, because if you're using something like large language models or other AI aggregators used in decision support systems, it just becomes a cover for indiscriminate targeting when you consider how inaccurate these models can be. Something this inaccurate also doesn't help you achieve your strategic and tactical military goals because you're wasting munitions and you're also unable to tell what you've really targeted and achieved.

Adam Wishart (32:16)

As a sort of historic example of how this is used, I interviewed three Israeli intelligence officers who had initially been called up in the late 2000 and tens, and then they turned back up in the days after October 7th. And in their point of view is that there was A demand for creating many targets in Gaza. And the only way that that could be achieved is if you turn to non human ways of doing it, into computers and these AI systems. And their argument was not only was there an inaccuracy, but it provided a kind of a blur or an excuse through which the Israeli military could purport to be hitting accurately many targets, but actually they were striking many civilians instead.

Venetia Rainey (32:58)

I want to get onto Israel and Gaza in a bit more depth shortly because you cover it and lots of interesting angles of that in your documentary, Adam, but just staying on Iran, on the very first day there was an attack on an IRGC naval base and taken out just next to it was a primary school in Minab in southern Iran. We don't know exactly why that school was hit. Was it outdated intelligence? But is it possible that that feeds into what you were talking about, the inaccuracy of AI when it provides multiple targets and that those targets are not properly verified by humans?

Heidi Klaff (33:33)

Yeah, I think it's actually a very good example of how AI models make it really easy to obscure accountability because the use of these systems makes it really difficult to distinguish if these civilian attacks were in fact deliberate due to intelligence failures or due to the lack of AI accuracy itself or a combination of all three. Because AI could have also been used to generate intelligence, but the black box and sort of scale and inaccurate nature of AI makes that really difficult to determine. Since we're talking about Iran here and also the US carrying out that strike, there's a really interesting investigation that showed in a recent strike on a civilian In Iraq in 2024, the US Central Command admitted to not knowing whether some strikes were in fact AI recommendations or not. And so if we have the Department of Defense or Department of War deliberately not recording when AI based decisions are being made, then it shows that AI is really being used to muddy the accountability here and makes it very difficult for us on purpose to be able to distinguish if this attack could have possibly been due to an AI accuracy or it could have been deliberate. And I think that's one of the Back to Adam's point. It takes that accountability away from the decision makers in the chain of command.

Venetia Rainey (34:42)

We should say that. I mean, your focus is on safety in AI, but you're far from the only person raising these sorts of concerns. In the week leading up to the Iran conflict, there was a huge row between Anthropic, the company that runs Claude, and the Pentagon. So their row was over the use of autonomous weapon systems and Anthropic at the time said and maintained still that its systems are not reliable enough for them to be used for those sorts of weapons yet. And it wanted more oversight and the Pentagon said no. Heidi, you worked at OpenAI when Dario split off to create Anthropic, the big AI tech divide that our listeners may or may not be aware of. But there it is. OpenAI said that it would take over those contracts for the Pentagon. Where are we at now? Who's providing the AI systems for the Pentagon? And is it even possible to replace Claude overnight when it's so deeply embedded in Maven?

Heidi Klaff (35:34)

I think currently we have an ongoing deal between Department of War and OpenAI, but Cloud is still powering Maven. It is very difficult to transition these systems overnight, if at all, because at the end of the day, Anthropic had a partnership with Palantir in which for the past couple of years they've worked on embedding these systems into Maven. This is not something that's going to happen overnight, even though Anthropic has been designated as a supply chain risk. Right. But at the same time, whether you're using OpenAI models or whether you using Anthropic models, you still have the accuracy issues. And these are the same types of models where OpenAI themselves had admitted to them inherently having hallucinations. And this is a problem that's not solvable. And in terms of Dario's, which is Anthropic CEO point of view on decision support systems, they are in an active partnership with Palantir and they are taking the stance against autonomous weapons systems currently because they believe that their systems are too unreliable. But the problem is decision support systems, which is how their model is being used, aren't really far off from being autonomous weapon system because of something that we have in the field called automation bias, which is this idea based on decades of research showing that humans often trust the recommendations of algorithms over other sources when approving them, often without corroborating if those recommendations by AI models were correct or not. And this is especially the case in military context, when operators typically only have a few few seconds to make determinations on whether or not an algorithm's output is correct. So this makes this distinction that Anthropic is making between decision support system and autonomous weapon systems pretty superficial in practice. So if Anthropic believes that their models aren't reliable enough for autonomous weapons systems, they're also not reliable enough for decision support systems. And I think OpenAI is going to have the same issue, regardless of the Contract. There's still the accuracy issue that remains.

Venetia Rainey (37:32)

Adam, you interviewed the head of Palantir's UK operations, is that right? Can you just tell us a bit about Palantir? Their names now come up quite a few times in this conversation. I'm sure our listeners will have seen them coming up as well. They're often described as controversial. Can you explain who they are and

Adam Wishart (37:45)

why they're often called controversial? They were started about 20 years ago. Peter Thiel was one of the founders. One of their first investors was the venture company of the CIA. And they did a lot of work in the early years in the national security apparatus of America and European countries post 9 11, trying to filter the signal from the data to find terrorists effectively. And subsequently they've been involved in ICE immigration in America in a very controversial way. And so they seem the kind of bete noir of polite society across Europe, which, given that Google, Amazon and Microsoft, for which many of us have subscriptions to, also are involved in many of these things, seemed slightly unfair in a way. In the early months of the Ukraine war, the Maven smart systems was used by NATO to give target assistance to the Ukrainians. And then shortly afterwards, Zelenskyy invited the CEO of Palantir, Alex Karp, and Louis Mosley, the head of UK Palantir, to Kyiv in order to buy the same system, because he wanted to have the same system that. That he thought had been so useful. And so the Ukrainians began to use Palantir's Data integration and AI system through 2022 and 23. And I think from all the reports, it seems to have been useful to them. And in the end, these are about human decisions. Humans are making decisions whether to use this AI system, whatever its inaccuracy, and humans are assessing whether or not the inaccuracy is something that we care about or not. And so one of the things I find interesting about delving into this is when you talk about target assistance against the Russians, people sort of shrug their shoulders and say, well, you know, it's an unjust war and we should fight it. When they say it's about the Palestinians, it causes everybody to raise concern. And that's because we, I think, as humans, make different decisions about each conflict.

Venetia Rainey (39:43)

Hi, did you want to come in on any of that? I see him nodding a lot. Yeah.

Heidi Klaff (39:47)

I mean, I think the way that I view it is that it doesn't matter who's at the receiving end of AI systems. There's international humanitarian law, which sets the boundaries of indiscriminatory targeting. It also focuses a lot on the predictability of these systems and keeping human decision making meaningful and not just something as like a rubber stamp. And so I think the way that I view it is from a general safety perspective, regardless of the conflict. But in terms of Palantir itself, yes, there are other companies that have provided cloud services for the militaries, or they might provide AI tools, but Palantir itself, there's been a concern about their power concentration and them gaining a lot of state secrets from the UK's overall reliance on them. And at the end of the day, they're essentially a private US company with its own interest. Right. And so when they have access to things like defense policing, nuclear and medical data from the British state and the ability to aggregate and analyze that data, Palantir can gain state secrets because it can then infer a lot of information, not just about the British population, but also about our national security through its use of AI and data analysis. You know, Adam gave the example of how Palantir systems are being used by ICE in the US to surveil its population. And I think if you collect data from different sources, including data brokers, location data, Internet happens, and now we have confidential private state data or data from the NHS to train or input these data sources into AI, you can now draw inferences on whom these individuals are and track them accordingly. And, you know, I've done research in the past where we've shown that AI being trained on a lot of our activity makes them inherently very well suitable for surveillance, which is sort of the other side of the coin to the military intelligence gathering capabilities of AI. And I think people fundamentally don't understand that despite that data technically belonging to the British state, the use of AI and data analysis by Palantir allows them to infer and gain insights in a way that wasn't possible by traditional software. That means that it remains in their possession. And so there's a question of also where that data is being stored and if other governments like the US would be able to have access to it. So I think people do have a right to be concerned.

Venetia Rainey (41:56)

Adam, in your documentary, you talk a lot about how Israel started to use AI to surveil the Palestinian population in the west bank and Gaza through a system called Lavender. Can you tell us a bit about that?

Adam Wishart (42:07)

Well, the historical starting point for my film, in a way, is around 2016. It's a significant moment, given that we've been talking about Maven, because it's the moment where Google had the contract to begin algorithmic warfare for the Department of Defense, as it was called then. And there was a worker rebellion. And Google gave back this contract and it was taken up by Palantir. But at the same time, these data experimentations in warfare and surveillance were happening in the place that sold itself as a technological superpower, Israel. And at the time, they had a problem. This was 40, 50 years after the occupation of the West Bank, 70 years after the Nakba and the foundation of the status of Israel. And in the west bank, the Israelis believed that they had a problem, which was that individual Palestinians were attacking Israelis with knives, screwdrivers, cars, in what they called the lone wolf in Tifada. And the problem for that is, in the old days, with a bit of human intelligence, you might have been able to disrupt a militia organization, but it's very difficult. As one of the whistleblowers witnesses in my film says, it's very difficult to work out who's going to wake up in the morning and decide to stab an Israeli. And so they decided they were going to use data analysis and effectively what we'd call an AI now in order to figure this out. And one of the advantages the Israeli state had in terms of the west bank is that they captured all of the telecommunications data in the west bank, so they could suck in all of the text message and voice calls into databases. They could make voice to text, and then intelligence analysts could interrogate the data to try and figure out who was going to be a terrorist that morning. And the first way they did that is to do it with keywords. You know, apparently people would send a text message to their girlfriend saying they would commit suicide at a checkpoint. The words checkpoint and suicide became red flags and would set off on alarms. And then later, they started using this pattern analysis. If we know that some person has all these attributes to do with their age and their gender and what employment they have in education, if those people are likely to be terrorists, then can you look in the full set of Palestinian data and figure out who that was? And what was interesting about it is, from the point of view of the Israeli state, it appeared as if this prediction policing was being successful. And it gave great power to the kind of techno evangelists of the Israeli state to pursue this idea that you could work out where the patterns were and who the people that were dreaming of violent rebellion were. But for human rights groups, it had a number of problems. The first is that the waves of terrorism in the west bank and Gaza have always ebbed and flowed. So perhaps the decline as it was from 19 to 21 was not because of predictive policing, but because people decided that that was no longer a modality, that was having an effect. And the second reason, it was the massive invasion of the privacy of millions of Palestinians and then also the accuracy, and particularly, most importantly, the Israeli state. Even if you believe the sort of national security ideas, the Israeli state invested a lot in technology and diminished their investment in human sources, which had been historically, over years, have been really central to the security of the Israeli state. And so by focusing on technology and not on human sources, they missed, you know, it's generally agreed that they missed the intelligence that was vital to have prevented the October 7th attack.

Venetia Rainey (45:26)

So October 7th, 2023, Hamas launches its brutal attack on Israel, kills more than 1,200 people. That sparks a war, a brutal war in which Israel bombs Gaza heavily and uses this system again to find more targets who might have been linked to October 7th or Hamas in some way. Can you tell us a bit about how that worked and how that generated. Generated more targets than the Israelis could ever possibly have found without AI?

Adam Wishart (45:51)

One of my whistleblowers said that at the beginning of the Gaza war in October 23, there were murmurings from the high ups that there just weren't enough targets. And the political impetus was to attack Gaza in a way, I think Netanyahu said, in a way that they have never known. And so they wanted lots of targets. And as the whistleblower said, the only way you could do that was non human methods. And so the non human methods was to build on this surveillance system that they had previously and to create what they called a target factory. They even issued press releases about their target factory. And the target factory was to generate loads of targets. And effectively there were a number of different systems, Lavender has previously mentioned in the press. It's not clear whether that is an AI system or a spreadsheet that is fed by AI systems. And these different systems worked in different ways. So some of them sucked up all the telecoms data that they had used a decade before, and you could work out associations between different people that they knew were a terrorist or had previous intelligence, was a terrorist. Who were they phoning? Were any of those terrorists? They had others where they analyzed the population according to attributes. What does a terrorist look like? Maybe these are terrorists and they had others to do with locations. And so as a consequence, they created many, many more targets than previously. In the first three weeks of the war, they had 15,000 targets. But ultimately, what the whistleblowers say is more than anything, these were human decisions about what was the accepted collateral damage. And so because of this desire to have an all out war in Gaza by the Israeli state, they made decisions to allow collateral damage. And that collateral damage at the beginning of the war was raised. It didn't matter if you killed 20 innocent civilians along with one supposed terrorist or actual terrorist, whatever it was. And the whistleblowers also make the allegation that for high level Hamas commanders, you could take out 300 innocent civilians. And that seemed like in the data, it seems that the Israeli state sort of confirms that there was a particular attack on December 2, 2023, that the Israeli Air Force announced that it took out a high level Hamas commander who'd contributed to the October 7 planning. And it's estimated that 300 civilians were taken out with him. I mean, it should be said that the IDF deny this. They say that they don't do AI targeting, that they just do decision assistance, that they act according to humanitarian law, that all decisions about targets are made in the presence of, with the expected civilian loss has to be in proportion to the military gain.

Venetia Rainey (48:35)

Yeah, I've got some stats up in front of me here. Within a month, the IDF said it hit 12,000 targets, within a year, 40,000 targets. And we do have relatively final death tolls now for the conflict. After years of denying the Gaza Health Ministry numbers, Israeli officials admitted earlier this year that they were broadly accurate. 70,000 killed, around 20,000 of whom were children. Heidi, what does this tell us about the safety of these systems when you're operating in a very dense environment where the military infrastructure is embedded within the civilian infrastructure?

Heidi Klaff (49:07)

Well, I think, as Adam pointed out, when you're creating these target profiles, you're implicating civilians without evidence or actions. Because these algorithms use statistical and probabilistic calculations from historical data to predict where future targets may be. And the emphasis here is that these are predictions, which does not necessarily reflect reality. A prediction might work and a recommendation engine telling you what movie you might like watching next. But you know, when human life is at stake, that, you know, the equation is very different here. And you're implicating a lot of civilians, especially when you're looking at these loosely defined parameters that expand to meet a specific target quota. So if, for example, you are against a specific militia, in this case Hamas, and you're looking to expand your targeting parameters, and also at the same time you also have these really high amount of casualties, you're willing to accept, as Adam was saying, you can then say, well, anyone in the. With a WhatsApp group, with a Hamas sort of militia, but it could have been someone's dentist, right? Like this is someone that could have not even known that this was someone who's part of a militia. Right. When you expand those parameters, you're inherently sort of blurring those lines yourself. Right. So it's not just about a military being embedded within a civilian population, it's just the nature of when you're creating these targeting profiles and you're in some sense expanding them to meet a superficial quota target. And this exactly goes back to the point I was talking about with speed, right? When you're looking at speed and you're sacrificing it above all else, you're going to sacrifice accuracy and you're going to also have casualties in the process. And I think when we look at the destruction in Gaza, I don't think anyone can say that the use of AI there was precise, given the wholesale destruction that we've seen across the Strip.

Adam Wishart (50:53)

One of my whistleblowers says that if you want an all out attack on Gaza and if you want to say that you're going to abide by humanitarian law, then what AI does is provide what he calls an excuse. And so it allows you to tell your military and your soldiers that this is all legitimate, whilst the underlying premise is not true.

Venetia Rainey (51:18)

So it's this idea of accountability and where does accountability lie? I want to take a step back. What do you think? Bigger picture and looking forward, how is AI going to change warfare? We're sort of just at the beginning of it all now. There were some claims in your documentary, Adam, that it will enable the west, broadly speaking, to be quicker than the enemy and that that's important from a military perspective, but there's also a suggestion that it will make wars more frequent and more brutal. Heidi, do you want to kick us off?

Heidi Klaff (51:46)

I'd actually argue that this isn't a new era of military strategy or technology. I think what we're seeing happening here is AI being used to evade accountability and also international humanitarian law. When you sort of keep in mind the abysmal accuracy rates that we talked about, right, 20, 50% that we've previously seen, the speed of large language models or other AI making these recommendations. And then you also see sort of these loosely defined parameters that are being given by the militaries that do accept a huge amount of casualties. How is this different from a high tech version of indiscriminate bombing? And I think we should then question How AI is being used to create loopholes in a legal system that is meant to prevent against the. This very thing happening.

Adam Wishart (52:29)

You know, for the evangelists of this technology, they say AI is going to prevent war through deterrence or make them quicker and less violent. And, you know, I think history doesn't bear that out. What AI seems to be doing and drones as well, you know, it seems to be, if anything, lowering the cost of beginning a war. And if you can lower the cost of, of beginning a war both politically and in terms of military technology, then when things cost less, they happen more frequently. And the uncertain world in which we live in might be bearing some of that out, sadly. But, you know, I think it's the very last word of the film are from a Ukrainian drone pilot who says that all wars bring technological innovation, and those people who can adapt it and use it faster are the ones who survive. And there's a sort of brute pragmatism by which the west is attacking this, I think, which means that there's definitely going to be more of it happening.

Venetia Rainey (53:31)

And we haven't even mentioned the fact that Claude was also used in the American attack on Venezuela to depose Nicolas Maduro. So I suppose that feeds into this idea that you can use AI to get things done more quickly, more efficiently and clearly Trump thoughts that Iran would be some kind of repeat of Venezuela, that it be quick and efficient, which it has. It hasn't been so far. What do you think needs to be done? We can't put this genie back in the bottle now. How do we regulate this going forward? Is it possible to be regulated?

Heidi Klaff (53:57)

I actually think that there's a lot of military standards out there, which is, you know, I used to work on the auditing of these types of systems. And military and defense standards, especially in the US Are some of the most stringent and rigorous standards there are in terms of understanding how to evaluate systems and their reliability. And typically, when we're talking about safety critical systems and defenses especially, we're typically thinking about something like 90% reliability. Right. Because if these systems fail, like I said, you're not meeting your strategic goals. If you have something that's indiscriminate and with 30% accuracy, you're wasting munitions. You're not sure if you're achieving your strategic goals. And so I would say actually, since the Post World War II era, we do have these very frameworks. But what's actually happening today is that through a lot of new contracts with the Department of War, we're seeing a fast Tracking of these large language models being used and deployed, where the companies are actually positioned to grade their own homework. We've seen several cuts to departments are meant to independently verify and validate these types of systems to see if they're fit for purpose. And we're instead seeing this new offshoot R and D program that kind of accelerate their use. And so you have these other companies, it's not just anthropic, right? You have companies like scale AI who are winning contracts to validate their very own systems that they're providing. And we should be questioning why that's even happening. And so I actually think that we don't need to invent the solutions, that we don't need to invent a new type of framework or new type of regulation. If we actually abide by what sort of been the norms that have been established in the past few decades, I think people will find that these systems aren't reliable. And I think choosing not to abide by them speaks more for the use of AI being used to evade accountability and use for cover for specific actions more than anything else.

Venetia Rainey (55:44)

Adam, any thoughts on that?

Adam Wishart (55:47)

Well, I think the scary thing is the speed of adoption and the insidious nature in which they become adopted. I think when I began this, I imagined there were going to be autonomous drones firing across the Ukrainian battlefield with embedded AI. And I think what really is the scary thing is that intelligence analysts might use a little bit of AI and then it might get passed onto a target room where people, they just assume it's another human target and then those targets but put into a drone with what Ukrainians call computer vision, which is another kind of AI. And so people talk about humans being in the loop, but it seems to me that humans may be in the loop, but they may be the empty vessels while the computers do the thinking. And I think that's the great worry.

Venetia Rainey (56:39)

Adam, you mentioned there, that was one of the things that I did find quite scary about your documentary that Ukraine was making these drones that fly themselves, choose a target and kill on their own without any sort of human intervention. What did you think about that?

Adam Wishart (56:52)

Well, the Ukrainians are very pragmatic. They sit in bunkers, you know, in this weird muddy 30 kilometer zone where no humans should safely go because drones can get to them. And the Russians have such have developed their electronic warfare sufficiently that you can no longer fly with radio signals, drones to targets. So what are the solutions? That one solution is fiber optic cables where you can keep the connection between the drone and the operator. But the Other solution is to plug in the coordinates of a tank, say, and then the AI can follow it. The trouble is that once the drone arrives 30 km the other side of the front line, and they see that the tank is a tractor rather than a tank, the operator is no longer in control. And so the concern I have is that humans aren't sufficiently aware of these consequences and that rapid uptake of AI is so huge that there's a sort of inbuilt technological determinism in which people think that this is absolutely the only way to go, rather than thinking that humans are making these decisions and they have the ability to make the good ones or the bad ones.

Heidi Klaff (58:08)

I think the example that Adam mentioned actually shows how a lot of military operators are actually unaware of how AI works, because if you're an AI expert, anyone could tell you that is what would happen, right? AI is unable to sort of keep context. And also it often just mimics the behavior of the training data that it's been fed. So in any sort of new situations, it doesn't know how to generalize outside of that. And I think there's a really big disconnect between military personnel and also the hype that they're being sold by these AI companies and how brittle these AI systems are. And so there's often a disconnect when evaluating these systems or making determinations if they should be deployed on the battlefield, especially when used in something as consequential as autonomous weapon systems. Because as an expert in AI and safety, I could have told anyone that that is ultimately would have been a result.

Venetia Rainey (59:00)

That was Heidi Hlaff, chief AI scientist at the AI Now Institute, and Adam Wishart, whose new documentary Click To Kill the AI War Machine is streaming now on Channel 4. That's all for today's episode of around the Latest. We'll be back again tomorrow. Until then, goodbye. Around the Latest is an original podcast from the Telegraph, created by David Knowles and hosted by me, Vinny Sharaney and Roland Olyphant. If you appreciated this podcast, please consider following around the latest on your preferred podcast app App. And if you have a moment, leave us a review as it helps others find the show. To stay on top of all of our news, subscribe to the Telegraph, sign up to our Dispatchers newsletter, or listen to our sister podcast, Ukraine the Latest. We're still on the same email address battlelinestelegraph.co.uk or you can contact us on X. You can find our handles in the show Notes. The producer is Peter Shevlin the executive producer is Louisa Wells.