This is an interview with Erik Reichborn-Kjennerud. Erik Reichborn-Kjennerud is a Senior Research Fellow at the Norwegian Institute of International Affairs (NUPI). His research sits at the critical intersection of war, technology, and political violence, exploring how contemporary warfare is shaped by military theory, digital technologies, and operational practices. Reichborn-Kjennerud’s book, The World According to Military Targeting (MIT Press, 2025), offers a critical investigation into how the operational logics of military targeting have come to shape a worldview defined by military supremacy and endless conflict. His interdisciplinary work draws on critical IR theories, Science and Technology Studies, and the history of science and technology.
DIFFRACTIONS: Could you provide a bit of foundation and background for people who are not familiar with your work? What ultimately drew you to study military targeting not exclusively as a task of identification and engagement, but also as an epistemic practice? For someone who sees targeting only as a logistical or tactical problem, how would you introduce concepts such as martial epistemology or world‑making?
Erik Reichborn-Kjennerud: The story is quite long. In the heyday of the so‑called war on terror, there was a lot of attention to drones and drone warfare, and also to some extent to military vision systems and what were known as signature strikes at that time. But that attention was mostly connected to the technical task of finding and killing. I am Norwegian; I live in Norway, a small country, and I happen to know a few people in the military, especially special operations forces who have been to Afghanistan. They talked about what was going on there in quite different ways than what was appearing in scholarly and media accounts focused on drones. They talked about targeting as a process. The whole idea of identifying or finding an enemy was a process that some people were excluded from but knew about. They introduced me to the American concept called F3EAD, pronounced “feed”, which is a targeting process the Americans had incorporated to hunt individuals in Afghanistan and Iraq.
To me, when they talked about this, it sounded much more like a process of knowledge production: figuring out what the world was, what the enemy was, and how to operationalise these understandings. That drew me to try to grasp what they were actually uncovering. So this is how I became interested in the history of targeting and how we ended up in this particular way of producing knowledge about the world. At the same time, I was very influenced by Antoine Bousquet’s excellent book, the Eye of War (2018), a historical account of the increasing convergence of perception and destruction. While his book was about the martial gaze, I guess one can say I was more interested in the martial mind, not how they see what they see, but how this seeing is already structured by certain ways of knowing the world.
There was also a couple of other foundational articles that steered me in a certain direction. One was by Astrid H.M. Nordin and Dan Öberg who wrote about war and targeting as a process – not explicitly an epistemic process, but a very procedural way of thinking within the military. The other was Peter Galison’s The Ontology of the Enemy: Norbert Wiener and the Cybernetic Vision that chronicles the different ways in which the enemy was made during WWII. Then I became interested in the history of science and science and technology studies, especially Paul Edwards’ The Closed World: Computers and the Politics of Discourse in Cold War America (1996) and Donald MacKenzie’s An Engine, Not a Camera: How Financial Models Shape Markets (2006), but also more general in questions about data and classification practices. All of this pushed me further toward the epistemologies of targeting: how targeting is not just a way to find and kill people, but a way of producing knowledge about the world, to create worlds that the military can operate in.
DIFF: Could you unpack the concept of martial epistemology a bit further?
ERK: Part of this for me is to draw on what is known as onto‑epistemology or the idea that whatever we know about the world is inseparable to how we produce knowledge about the world, or what we can call world-making. Drawing on historical epistemology, people’s understanding of the world is connected to the means by which they understand the world. Consider Galileo’s telescope: he understood the world through the telescope, which was a novelty at that time because he could actually observe what was going on in space. But this was all connected to the means, the telescope, to observe the world. Different vocations understand the world through different means, and therefore understand it in different ways. That is the epistemological dimension. I focus not on the question of what is, which is ontology – but on how what is, is inseparable to the fundamental question of how we produce knowledge about that. In that sense, the world becomes real through these practices. For me, targeting is one of these epistemic practices that brings the world into being. And then, of course, targeting is also a way to destroy that very world that they have produced. So, the martial epistemology is a violent operational epistemology.
DIFF: I would like to turn to your book The World According to Military Targeting, specifically the chapter “Hamlets, Humans, and Worms: Computing the Environment in Vietnam.” Vietnam served as a crucial testing ground and milestone for automated sensor technologies. Could you introduce us to three key programmes from that era: the Hamlet Evaluation System, the Phoenix Program, and the Igloo White sensor field? More importantly, why do you identify these as essential case studies for understanding targeting – not targeting understood as a crosshair or a tactical strike, but as a practice of profiling people, and of finding, discovering, or even producing threats?
ERK: When I started writing about this, there were already excellent case studies on these three systems, which had pointed out their historical importance to the present — particularly the global war on terror. What I wanted to show, apart from highlighting the specificities of how these different systems produce knowledge and their links to targeting, was that together these three systems demonstrated how one attempted to hunt people in Afghanistan and Iraq. The reason is that to find particular people, you need to compute the whole environment: you need to know the population and the terrain. Or, this is how the martial apparatus thought of it. What they were actually doing was to produce the enemy. Together, these make up what could be considered the totality of the enemy in Vietnam. Previous cases had shown, for instance, that the Phoenix Program – a man‑hunting program in Vietnam had specific connotations to what was happening in the war on terror.
Igloo White, a sensor‑based surveillance system connected to early drones, was an automated bombing system that also had clear connotations to the present. But together, these systems show what I would call the martial dream of knowing it all. In the Vietnam War we see a real interest in, and a lot of funding towards, computing everything they could come across, producing data about everything they believed was important. The US military believed that to win the war they had to kill people, but to do so, they needed to know the population, they needed to know individuals, and they needed to figure out a way to track movement, and preferably automate all of this. That was a large part of what they tried to do in Vietnam. Of course, these systems were complete failures, laying waste to an entire country, its environment, and society, but their dreams and their fantasies live on today. We clearly see their legacies in contemporary wars, especially in the war on terror, but also in the enduring dreams of transparent battlespaces and the belief that one can calculate and predict oneself to victory.
DIFF: Would you say Igloo White was one of the first iterations of a sensor‑driven battlefield? Robert McNamara declared Vietnam the first techno‑war or technical war. Do you regard it the same way?
ERK: Previously there were many sensors out there, but what Igloo White did was to automate a lot of the process ‘from sensors to shooters’ as the military likes to put it today. Tracking through sensors was a novelty at the time, and you could also estimate where to bomb based on those sensor feeds. It was highly experimental and a failure, but it showed the military and martial apparatus that this could be done in some sense, and it fueled their fantasies about what was possible. The lasting legacy is the belief that we just need more sensors to produce data, more processing capacity to process the data, more network capacity to link everything together, and also more shooters.
This is linked to what is now known as the kill chain and the dream of truncating the kill chain into seconds rather than minutes. But let me qualify that statement a bit. Obviously, the computer, sensors and data were crucial for the US military war effort in Vietnam, but there were at the same time a number of other things going on in the US that contributed to sensor-driven battlefields and data-driven war, like the Semi-Automated Ground Environment (SAGE) and the Bombing Encyclopedia of the World. What connects these and the various programs in Vietnam is that they all depended on turning the world into data, or datafication, and that data into operational knowledge. So while Vietnam tells part of this story, it is also a larger history of making the world machine readable.
DIFF: I was also struck by the constellation of thinkers you engage with: Aradau, Chamayou, Deleuze, Foucault, Massumi — and how you link targeting to questions of the norm, the baseline, and anomalous behaviour. Could you particularly spell out how you work with Foucault, Deleuze and Chamayou and against them? As Foucault could be read and as you explicate, in the disciplinary society, the norm operates as an ideal type: a statistical distribution around a predefined standard, against which deviations are measured and corrected. In the control society, as Deleuze describes it, norms become continuous modulations, ever‑adjusting algorithms, where the subject is never fully inside or outside, only perpetually in variation. Your work seems to point toward a third register, a ‘normativity without a norm.’ As I understand, the system does not start with a pre‑existing signature of enmity, nor even with a continuous modulation of a known variable. Instead, it generates a baseline from the data itself, and the anomalous is defined purely as whatever deviates from that immanent, self‑updating distribution.
ERK: One key thing during the war on terror was that the martial apparatus were creating what were known as “signatures of enmity.” This could be considered an ideal type, but not a statistical norm in the Foucaldian sense. They had created a particular signature of what the enemy was supposed to be. What a terrorist is according to them, and they were looking for that very signature to kill and capture those people, and the systems set up to find these signatures found many, often too many, so-called false positives. But they also had, they believed, the problem of what Rumsfeld would call the “unknown unknowns” people that they could not force beyond the threshold of detection of the martial sensorium. That meant the signature didn’t fit. In a short article called Patterns of Life: A Very Short History of Schematic Bodies Grégoire Chamayou points out that a technique called activity‑based intelligence (ABI) was a way for the US military to get at the unknown unknowns.
Rather than starting with a signature you are looking for, going into an area and looking for bad people so to speak, here you have a technique that does not base its hunting on a particular signature or a pre-defined Foucauldian norm of what is good or not. It is not a predefined thing or behaviour that defines enmity. Instead, enmity or the enemy emerges through the computational technique of anomaly detection – a technique that is basically about producing a normal baseline of everything that happens, and anything that sticks out (the anomaly) from that normal baseline is, by definition, a problem — suspicious or a threat. You start by compiling all the data you can. You are interested in the normal operation or data operation of a particular area, and anything that sticks out is given a red flag and you go after it. Rather than the Foucauldian norm based on a statistical distribution, this is interested in the anomalies that come from looking at the baseline. It produces enormous numbers of anomalies, of course, because people do not behave the same every day.
Chamayou theorises this as the “targeted society” – though he is talking about this targeting technique being applied everywhere, from marketing to war. In the book I am more interested in how ABI is part of a total encompassing system that tracks not only particular signatures but also everything that is not the signature. Rather than being a control measure or a security apparatus, I think about this more and more in terms of domination. The military is more interested in dominating than in controlling, being able to target whomever and whatever rather than surveilling whole populations. But to be able to target individual targets you have to process everything. And this is why anomalies and activity‑based intelligence are part of a larger structure through which the U.S. military apparatus was hunting people during the war on terror and still does. But it does not function alone.
DIFF: Could you expand on how it does not function alone?
ERK: Partly it is about particular individuals that the U.S. military apparatus knows the identity of, the bin Ladens or Saddam Husseins of the world, they try to hunt down, they know who they are, they just don’t know where they are. Then you have all the people they have no clue about, but they pop up when they hit the thresholds of the signatures that the U.S. military has made. A signature might be anything: proximity to an area under surveillance, your phone having called another phone connected to a network of terrorists, or something more sophisticated. For example, the Skynet program was about trying to seek out couriers, especially in Afghanistan. Couriers were believed to be the key node, because many terrorists had turned off their phones and gone off the grid.
They believed couriers would lead them directly to high‑level individuals. They didn’t know who the couriers were, but they created a signature based on seven known couriers. They took their behavioral patterns and fed them into a machine learning system based on a random forest algorithm, then tracked the digital ecosystem to find people who behaved in similar ways. In that sense, you could easily become a signature. People behave in much the same way: they need to eat, go to work, so signature-hunting would then find a lot of people. Anomalies are then part of another part of this. The signatures could make everyone look like a terrorist if tweaked incorrectly. There was also a paranoia that there was always someone they couldn’t find, and activity‑based intelligence tried to solve the problem of unknown unknowns. But of course to find what sticks out, you have to know everything else. It becomes a grandiose idea that you can and have to know it all – that was Keith Alexander, former head of the NSA’s slogan: “Collect it all, tag it, store it. . . . And whatever it is you want, you go searching for it”. So, ABI does not work alone, it is one part of the US martial apparatus dream of filling in the two-by-two matrix of known knowns, known unknowns, unknowns knowns, and unknown unknowns of the world.
DIFF: You analyse the F3EAD targeting cycle (Find, Fix, Finish, Exploit, Analyze, Disseminate) as a “self-referential feedback loop” in which raids are not merely lethal acts but epistemic probes – they “perturb” the enemy system to force it into visibility. As you note, citing Eyal Weizman’s account of an Israeli general: “Raids are a tool of research . . . they provoke the enemy to reveal its organization. . . most relevant intelligence is not gathered as the basis upon which attacks are conducted, but attacks become themselves modes of producing knowledge about the enemy’s system.” How does this collapse of violence and inquiry blur the line between warfare and a macabre form of field experimentation on living populations? What are the unspoken epistemic criteria for a “good” raid when its primary output is not territory taken but intelligence generated?
ERK: The whole idea behind the F3EAD (“feed”) targeting cycle is that it is supposed to be a self‑referential feedback loop in which the most important part is not the actual killing or capturing of an individual, but the exploitation of data that you can get from that individual and whatever can be found at their location. Raids, in the old‑fashioned military way, is about killing your way to victory. But in that process, intelligence would also stop; it would be dead ends. Drone strikes, in particular, were dead ends because they could not interrogate or torture the individual, and everything at the location was blown up. In this sense, special operations forces were more interested in capturing the individual for so-called enhanced interrogation, to see what the person knew, and also in finding phones, computers, or even regular notes. The idea was that this data could feed back into the F3EAD cycle, producing new individuals to go after.
They did this for a long time, killed and captured a lot of people. Probing or forcing the enemy to act, has a long history. In the First World War, you would send people out of the trenches to get the enemy to fire at them; when you drew fire, you could shoot back.
That is a simple way of thinking about it. But the Israelis, and U.S. special operations forces who had long collaborated with them, knew about how the Israelis thought about this. This type of probing was more about not drawing fire but about collecting data.
In the book, I quote a conversation between Michael Hayden, then head of the NSA and later head of the CIA, and the head of Joint Special Operations Command. Hayden was complaining about not having enough data to conduct raids, and Michael Hayden said plainly: if you give me a little action, I will give you more data. The idea is that going on the offensive gives you data. Raids were used not just to gather intelligence from captures or kills, but to force the presumed enemy to act. This reverses the classic OODA loop – observe, orient, decide, act. Here you start with the act, with action, to then be able to observe and orient.
You fight for information rather than with information. This has connotations back to Vietnam. There were ideas in Vietnam that to uncover the unknown enemy, you needed to poke at the enemy’s network to produce agitation, and then you could gather more data. The problem with this data‑driven warfare is: where does it stop? Why do you think violence will solve the problem? It only feeds the self‑referential feedback loop rather than producing anything of value. It is a macabre, violent way of producing knowledge. I think we can also think about the ongoing slaughter in Gaza and Lebanon as a way of testing these things, perhaps not so much probing the enemy as probing the international community and their allies to see how far they can actually take this.
DIFF: That is an interesting point. Reading about Israel’s assaults and invasion of Lebanon, they are not necessarily striking Hezbollah targets but destroying residential buildings and non‑military infrastructure. Hezbollah has been effective with drones against Israeli assets. I thought the IDF would be more formidable. The case in point about probing the international community’s reaction is interesting.
ERK: I think they are pushing as far as they can. With Trump also moving in to say they should not retaliate against Iran – the Israelis probe all the time. When there is no reaction, Gaza happens. Utter devastation. And with Iran: the administration admits they did not necessarily strike all military assets. Iran’s inventories of ballistic missiles, drones, interceptors, and launchers remain largely intact. In the first weeks, Trump kept declaring they had destroyed all military assets, the navy, obliterated them. Then it turns out most assets are still intact. So, the question of probing in this setting is not as much about finding where military assets are, but about figuring out the enemy’s and international community’s baseline for reaction and escalation. It is a continuous dance macabre or theater of horrors between adversaries, and the result in Gaza is genocide.
In the book I talk about Brian Massumi’s concept of “incitatory power” where the idea is to force the enemy to take shape. You believe the enemy is out there but has not taken shape, so you force it into being. By provoking or probing, you force the enemy to react in a certain way where you are ready to activate violence if need be. You are forcing not only individuals to flesh out, but the whole idea of the enemy to take formation and become the enemy you want it to be.
Partly, this kind of thinking, which is related to the previous question, is about driving signatures above so-called detection thresholds, fighting for rather than with information, but it is also linked back to the self-referential feedback loops of the F3EAD. The major problem with this is that it has no telos, no end state in mind, no idea of victory. It is only warfare unleashed to experiment on the world and see what happens. So I see the so-called war on terror and the current horrors in the middle east as part of a longer trend about what war is becoming – away from end-states, not a tool for population governance, but an experimental tool on the world in which the old rules are dying and genocidal acts are back. We tend to see war as a phenomenon bounded in space and time, but it is not. War is generative and can be made into a political tool, but its nature is about serving its own ends, seeking to perpetuate itself.
The old targeting regime derived from the 1930s systems theory, that was largely active throughout WWII and the Cold War and is still today – industrial web theory or systems warfare – had a logical and finite end to war. It was a kind of thinking that predicted a certain collapse of the enemy by taking out particular infrastructural targets. In this sense, if you destroyed all the designated targets, the enemy would collapse. At least this was the thinking. Once you had run out of targets the war would be over. Today, it is different.
I think we are now at a moment in time where what Dan Öberg would call “transgressive creativity” have met up with advanced computing – machine learning and so-called artificial intelligence – to accelerate and perpetuate warfare in a world that is seemingly out of control and in which the only solution is the martial apparatus and their supposed ability to not only bring order but to generate new types of order, control, and domination through never-ending violence. Computing also runs on ideas of feed-back loops demanding more and more data, and they will always produce an output. Together, solution-driven military design thinking and AI is a scary and potent mix that will accelerate and make war more lethal, with devastating consequences.
DIFF: You describe the martial fantasy of a “Map of the World”, a living digital twin that lets an operator act in and on the world without needing to understand why something happens, only how it can be made actionable and targetable. How does this aspiration to turn all reality into a targetable interface transform our relation to the world itself? What is lost when the cognitive goal shifts from making sense of a complex, ambiguous situation to operational grasping rendering everything into a manipulable surface? Can a purely operational ontology ever be wrong in a meaningful sense, or only inefficient?
ERK: The phrase “map of the world” surfaced around 2016–2018 within the geospatial intelligence community. They wanted to make a digital replica of the entire world – not only a grandiose fantasy but also a potentially lethal system for destroying objects by connecting all the known-un/knowns and un/known-knowns in a single model. For the military, it was about taking everything they know of interest and putting it into a large database (not really a map, but a massive database) where operators could search and test. It would be connected to a live feed, so you would see what was going on with these typed objects at the same time. The thing is, it does not give you any idea why something is happening; it only produces knowledge about what is happening, and a very particular version of the world. It excludes much more than it includes.
If you are interested in hitting objects, you might not be attuned to why. This has been a problem for the U.S. military for a long time – the idea that you can just hit a lot of targets and then win, because time and time again the idea that war is fundamentally a data problem has failed to provide the US with victories. In the map of the world, you would have all the known knowns: infrastructure, installations, static military assets (e.g., Iran’s oil refineries), everything a modern society needs to function. You would also put in moving objects – cars, people. You can query the system and find anomalies.
You can use it for activity‑based intelligence, feed it different sensor feeds. In fantasy, this would make up a living digital twin of the entire world – at least what is interesting to the martial apparatus. The idea of digital twins did not grow out of the martial world itself, but it is interesting to see this phantasmatic idea of knowing everything and containing it in a single database that you can access, analyse, and use to calculate your way to victory. Companies like Palantir are trying to make this aspiration a reality, explicitly calling their systems digital twins. With all the promises and fantasies around AI, large language models, machine vision systems, they are going full throttle.
DIFF: You mentioned that your current work focuses on large language models and retrieval‑augmented generation (RAG). I am also interested in companies like Palantir and Scale AI. NATO is using Palantir’s Gotham and the Maven Smart System. How does your work engage with these private, proprietary companies and the shift in the model? This connects to a point made by Elke Schwarz in a presentation at your institute. She observed that winning and survival are much more easily achieved in a business context than in an actual war, because war always entails a complex set of sociopolitical dimensions. That suggests a sliding scale: at one end, classical military victory that is elusive, messy, politically contingent. On the other hand, the concern regarding quarterly profit margins that is more measurable and far more attainable. As these private companies become integral to targeting and battlespace management, how do we theorise the conflation of these two registers?
ERK: One observation: with metric‑based victory talk out of the U.S. administration, counting targets hit, strikes made, the body‑bag world from Vietnam – they have also started talking about how many “tokens” they have used in planning against Iran. [Daily usage hit approximately 20 billion tokens during Operation Epic Fury]. I have no doubt someone is making tons of money, not just missile manufacturers but those who live on tokens. There is a lot of interest in commercial AI companies. Palantir is significant, but it is not really an AI company; it does not make what anyone would call AI, although what AI might be and how it is used as a commercial slogan are two different things. It is a software company interested in streamlining your organization from data to what they call logic to action. Palantir has major contracts within the U.S. Department of Defense and the different branches, and increasingly within NATO, especially through the Maven Smart System project.
What is interesting about Palantir is that it forces us to think about data again, not just large language models, and this is the focus of my current work with the brilliant Lucy Suchman. So what is it that Palantir is actually doing? Palantir goes into your organization — a hospital, a supply chain company, the military — and creates semantic objects from your data. By doing so, they define what your organization sees and what your organization can know. They use other AI companies like Anthropic or OpenAI, putting their LLMs on top of this. But in defining what the data is and the objects within data, they can, for example, go into the military and define object types like “military vehicle” and sub‑objects like “tank” or “lorry”. They make a whole semantic ontology. They also create people as objects – “person of interest” – with properties and automatic actions tied to them. They add an action layer with sensor feeds that feed into these objects; the objects develop based on live sensor feeds. If an object reaches certain thresholds, Palantir has baked‑in workflow triggers that can elevate a person of interest to a high‑value target without anyone noticing – you just see it in the system.
Palantir has created something quite ingenious from a business point of view. They are not selling a super‑intelligent AI; they have simply decided what your organization can see, know and thus what the world is for your organization. In the Lord of the Rings world, Palantír is the seeing stone, what they have now made through their so-called Ontology is a ring to rule them all. Now, once they control your data – they don’t own it, but they control what the data is, how it is structured, and what you can see – that is quite powerful. I don’t think we have understood that yet. We have been too concerned with LLMs rather than what Palantir is actually up to. In addition, they have become so powerful that if you are another company and want to move into NATO or U.S. DoD systems, you need to work with the Palantir ontology.
That means every software developed for Western military power must use Palantir’s ontology, further sedimenting Palantir as the foundational company. It is not only a problem that these companies define objects and their relationships, but also, and I think this is important to remember, that to compute something, to process data, you need to make the world machine readable. So what I am getting at, is that the question of ontological categories – of what is of this world – is already a problem long before the companies take over this task. IBM and their machines were in large part running the Vietnam war, for instance, forcing any knowledge generation to be compliant with how their machines worked. And then of course, there is a lot of problems with privatising intelligence, security and warfare, but for me that is a slightly different question from how we ended up with a world where we think that reality is simply out there, ready to be collected with the right epistemic tools and practices. This is the fundamental problem, and that goes beyond the commercialization or privatization of these techniques.
DIFF: In your current work, how do you assess the actual epistemic vulnerabilities of RAG for targeting? Does grounding an LLM in a classified database simply displace the problem of hallucination into the problem of database ontology – i.e., what counts as authoritative, what is omitted, and how the data were labelled in the first place?
ERK: It partly connects to Palantir. We talk about how stupid LLMs are, how much they hallucinate, how they are not connected to the data people need. The military has been aware of this for quite a long time. First, they have fine‑tuned LLMs on military language – targeting in a military setting is not the same as in a marketing setting, so they need the LLMs to understand their lingo. RAG itself is a way to bypass many of the problems with LLMs. Hallucination is largely a product of how LLMs store the data they are trained on to best generate text. RAG works as follows: you fine‑tune an LLM, but you force it not to use its own memory. Instead, it looks into a corpus of documents you have created – that is the retrieval part, or R in RAG. You prompt the LLM, and rather than generating from its own memory, it looks into the curated archive or library. This is a massive undertaking, but for the military, the corpus includes classified materials, their own databases, etc. For operational planning, they get outputs closer to what they think is right. But it does not solve the problem.
The LLM only looks at what the military already knows – it functions more like a search engine than something super‑intelligent. There has been a lot of critique within the military about this because many think it removes operational art as a creative activity. RAG produces outputs consistent with a structured way of doing operational planning – courses of action that look good but may lack the tacit knowledge of seasoned commanders and may not produce creativity.
It is very much designed around the highly structured way of thinking about operational planning, which the military has been trying to get out of for a long time. The martial apparatus is in this sense back to basics. The Ontology is a bit different, as discussed previously. While RAG systems are grounded in what militaries already know, Ontologies are grounded in what can exist, and thus also what can be known. But fundamentally, they both operate within martial epistemological structures and processes such as the operational planning processes and targeting processes. In many ways, the introduction of so-called AI into militaries, is not transforming or revolutionizing warfare, as its proponents claim, but rather accelerating and speeding up already existing martial processes.
DIFF: What are your thoughts on the emerging discourse around “world models”? Drones have long been trained within game engines, and the U.S. military has maintained synthetic training environments for decades. However, world models are now being pitched as something qualitatively different – not merely digital twins, but systems that generate causal, predictive simulation layers capable of reasoning about counterfactuals and anticipating emergent behaviors. In what ways, if at all, could world models transform targeting and the battlefield beyond what existing simulation and digital twin technologies already enable?
ERK: When it comes to world models, part of the question is how you define it and whether you believe in it. If we follow the famous computer scientist Yann LeCun’s argument around world models, it seems to me that this is all about making machines able to adapt to new domains and tasks, which humans are supposedly good at. Personally, I do not believe in it. A human does not have a world model either – we are all onto-epistemologically trained people, so what we think is real is always a product of how we produce knowledge of the world. The problem with current LLMs is that they are quite good at being flexible in generating text – you can ask them anything and get a good guess answer back.
But they do not really understand things, and as we all know they make things up due to how their memory based on training data is stored. In addition, they are just working with text or images, and the world is so much more than that. I do not buy the idea that world models will change everything. What we see is a lot of technical solutions to technical problems. We have reduced war to a data management problem – a long history of Western military thinking, especially among the technophile world. LLMs, RAGs, and computational ontologies are technical solutions to technical problems. They do not solve the so-called fog of war, and they do not lead you to victory. They dupe you into thinking you can compute or accurately define your way to victory. But war is generative.
A lot of things happen that computers and computing systems do not understand because they only compute what we already know or what we have set them up to see and understand. They do not plug into the generative character of war itself. There is a massive economy driving the current craze, which is dangerous. We have seen this in the bombardment of Iran – the speed and scale of destruction are unprecedented. LLMs and RAG are speeding up already known military processes. It is nothing more than acceleration of lethality which will likely speed up with new tools. So-called ‘world models’ will only be able to predict from data, (and all data is per definition historic) so if history always repeats itself down to the tiniest thing of this world then yes, they will be able to generate causal predictive layers capable of counterfactuals and anticipating emergent behavior. However, there is always difference in repetition, history does not repeat itself any more than humans do.
DIFF: Do you see possibilities for counter‑conduct or counter‑epistemic action against these targeting regimes? Where might points of intervention lie – whether in material practices such as camouflage or hiding, or in more systemic efforts to unmake or dismantle these architectures? Or does the challenge require a more fundamental shift in perspective: not simply evading martial epistemology, but critically understanding, exposing, and ultimately disassembling its internal processes, ontologies, and logics of world‑making?
ERK: We need a multi-pronged approach. There are different avenues. For starters, we need to better understand what is going on – untangle, unpack, criticise not only the Palantirs of this world but also broader militarism. We need to be much more interdisciplinary. Social scientists and humanities scholars need to talk much more with computer scientists interested in debunking what these companies are actually doing. Any disruption must be connected to deep knowledge about these systems – I don’t think we can just say that these are bad. We have to be able to point out the specificities of how they are bad. On the other hand, this is a massive economic undertaking, and the whole AI hype is so big that it is difficult to address.
There is a lot of activist work pointing out the problems with LLMs, with computing in general in society, with big tech deciding things – that is great, and we need more of that. There is also critique from within the military, which can help flesh out what is problematic. But ultimately, it all comes down to understanding the role of war in our societies. I find it extremely disheartening that Israel has been allowed to do what it has done in Gaza, Iran, Lebanon, Syria, and Yemen without much more international condemnation. Without that, it becomes difficult to see war as anything other than a political tool. Trump, who claimed to be a peace lover, has been conducting military operations and bombarding Iran for a long time. We need to dislodge that sense of war’s role in our societies — the militarism. War has a hold on politicians; it seems to be a good political tool for some, although history shows that most American wars have not gone the way they hoped. That is part of the book — it tells a story of failures. Targeting never succeeds. The appeal is in computing your way to an easy victory. Scientists love it, companies love it — technical solutions to the problem of war. But for me, the problem of war is something completely different. It is something we need to overcome and problematise as a political strategy and a solution to things. These companies sell a guide to victory, capitalise on it, and make money. That is where we need to intervene.
REFERENCES
Bousquet, A. (2018). The eye of war: Military perception from the telescope to the drone. University of Minnesota Press.
Chamayou, G. (2014, December 4). Patterns of life: A very short history of schematic bodies. The Funambulist Papers, (57). https://thefunambulist.net/editorials/the-funambulist-papers-57-schematic-bodies-notes-on-a-patterns-genealogy-by-gregoire-chamayou
Chamayou, G. (2015). Oceanic enemy: A brief philosophical history of the NSA. Radical Philosophy, (191), 2–9.
Edwards, P. N. (1996). The closed world: Computers and the politics of discourse in Cold War America. The MIT Press.
Faint, C., & Harris, M. (2012, January 31). F3EAD: Ops/intel fusion “feeds” the SOF targeting process. Small Wars Journal. https://smallwarsjournal.com/jrnl/art/f3ead-opsintel-fusion-feeds-the-sof-targeting-process
Galison, P. (1994). The ontology of the enemy: Norbert Wiener and the cybernetic vision. Critical Inquiry, 21(1), 228–266. https://doi.org/10.1086/448747
Greenwald, G. (2013). ‘Inside the Mind of NSA Chief Gen Keith Alexander‘ The Guardian (15 September. 2013)
MacKenzie, D. (2006). An engine, not a camera: How financial models shape markets. The MIT Press.
Massumi, B. (2007). Potential politics and the primacy of preemption. Theory & Event, 10(2). https://muse.jhu.edu/article/218091
Massumi, B. (2010). Perception attack: Brief on war time. Theory & Event, 13(3).
Massumi, B. (2015). Ontopower: War, Powers, and the State of Perception. Duke University Press.
Nordin, A. H. M., & Öberg, D. (2015). Targeting the Ontology of War: From Clausewitz to Baudrillard. Millennium: Journal of International Studies, 43(2), 392-410.
Reichborn‑Kjennerud, E. (2025). The World According to Military Targeting. MIT Press.