Abstract
AI is widely viewed as a technology of distinction and personalization. This article challenges this view by looking closely at the Israeli intelligence's use of AI in the Gaza War. Why was AI needed to produce what seems like indiscriminate mass killing of civilians and destruction of whole neighborhoods? Solving the paradox requires understanding technology within the legal, structural, cultural, and moral contexts in which it is embedded. AI was used to dramatically accelerate target production, since the Israeli military has adopted the “lethalness” doctrine (aspiring to maximize killing) but had also embedded international humanitarian law (which requires distinction between military targets and civilians) within its organizational structures, with lawyers in the kill-chain. AI was needed not to personalize treatment but to justify uniform treatment (bombing) by creating personalized justifications. It legitimated mass killing and destruction by automatically fusing and analyzing data to transform thousands of individuals and buildings into “legitimate targets” with individual probability scores (this deviated from traditional intelligence epistemology, requiring cultural work to overcome resistance). Each target was attacked for different reasons, linked to Hamas based on unique data pieces (crafted into unique stories by human officers). This legitimation of mass killing could occur even if the system were error-free. The critical study of AI must then move its agenda beyond errors and bias. AI's social impacts are revealed as more complex than often assumed and not inherently individuating: AI's core affordance of distinction may foster, simultaneously and at different levels, distinction and indiscrimination, individuation and de-individuation.
Keywords
How does the introduction of big data analytics and so-called artificial intelligence change social life? A very common answer to this important question is the personalization or “automated singularization” (Reckwitz, 2020) theory. In the big data society, the argument goes, individuals are no longer treated (and discriminated) as members of pre-existing social groups; instead, thanks to the intensification of surveillance and datafication, each individual is treated differently, based on one's documented past conduct and patterns in big data that grant conduct meaning. This meta-narrative (discussed in detail below) has been convincingly deployed by multiple scholars in sundry fields, including marketing, medicine, pricing, social media, policing, streaming, and warfare. This article uses a case study—the aerial phase that opened Israel's war against Gaza—to challenge it. This case presents an apparent paradox: the introduction of sophisticated big data analytics and AI technologies, promising fine-grained distinction and precision, ultimately produced what seems like indiscriminate mass killing and destruction. How could the introduction of technologies of distinction result in blurring international humanitarian law's (henceforth, IHL) core “principle of distinction” between legitimate military targets and protected civilians and civilian infrastructure?
On October 7, 2023, Hamas commando (“Nukhba”) units invaded Israel, murdered more than 810 civilians in border-adjacent Kibbutzim and at a nature party, raped women, and took 251 hostages (225 civilians and 26 soldiers). Consequently, Israel declared war on Gaza, officially aimed at releasing the hostages and ending Hamas rule in the Gaza Strip. More than 71,000 Palestinians had been killed according to both official Palestinian records and Israeli army estimations (Kubovich and Hasson, 2026), of which nearly 20% assumed to be militants and nearly half women and minors; thousands of others are missing and yet others died from disease and malnutrition; and nearly 2 million lost their homes as most residential buildings in Gaza were destroyed or severely damaged. The scale of civilian deaths prompted International Court of Justice proceedings against Israel for alleged breaches of the Genocide Convention. The damage to residential buildings and civil infrastructure, including schools, roads, hospitals, water facilities and bakeries, amounts to urbicide (Coward, 2009) or domicide (Rajagopal, 2022), systematic destruction of urban material environment that renders it uninhabitable and forms of social life relying on it unfeasible.
Killing and destruction occurred under various circumstances: during the ground maneuver; in attempts to hit Hamas militants moving in the field during combat; as revenge acts initiated by field units without formal authorization; and later (in summer 2025) in shootings at civilians trying to secure food and systematic demolition of neighborhoods by bulldozers. However, a large share of the killing and destruction was of targets identified by the Israeli intelligence through unprecedented use of sophisticated digital technologies. This article focuses on these airstrikes, particularly on the war's first phase (8–27 October 2023), before major ground incursions. In this phase alone, over 4300 Palestinian women and children under 17 were killed, constituting 64% of all fatalities, and entire neighborhoods were destroyed, evoking images of WWII carpet bombings. 1
This is the October paradox: the Israeli military boasted that target selection was done using the most sophisticated intelligence AI and data analytics tools to identify human and infrastructure targets linked to Hamas, yet the result of this unprecedented investment in sophisticated targeting gives the impression of indiscriminate destruction and killing. This paradox—thousands of civilians killed by an allegedly precise AI system—caught anthropologist Lucy Suchman's attention, who told NPR, “It appears to be an attack aimed at maximum devastation of the Gaza Strip. [If the AI system is really working as claimed by Israel's military], how do you explain that?” (Brumfiel, 2023). Similarly, Gusterson (2024) noted that the very existence of AI systems “suggests a concern with discrimination in targeting that seems at odds with the scale of civilian casualties.” If the aim is indiscriminate killing, why is such advanced technology required?
Solving this paradox may help refine the singularization narrative. Below I show that paradoxically, under certain conditions, AI may be used for indiscriminate mass killing not despite but because of its capacity for distinction: singularizing the justifications for indiscriminate killing, it produces distinction for indiscrimination. This happens, since AI's social effects are not predetermined by technological features but informed by the legal, structural, cultural, political and moral contexts in which technology is embedded. Particularly, indiscriminate killing and destruction stem from the (not-merely-)social construction of “targets.” In a time when IHL and military ethical and professional norms require aiming only at legitimate “targets,” big data's effect on war is not necessarily reducing mass killing and destruction by finding the needle in the haystack, as usually claimed. 2 In our case, it is the very opposite: transforming hay into needles, “incriminating” nearly every residential building by crafting a singular narrative uniquely associating it with the enemy, thus legitimizing indiscriminate destruction. Big data and AI are thus instrumental in enabling indiscriminate damage in the era of humanitized violence (Bonds, 2019). More broadly, I suggest that these technologies are not inherently individuating: their core affordance, their capacity for distinction, may eventually foster, simultaneously and at different levels, both singularization and generalization, distinction and indiscrimination. This necessitates rethinking the social meaning of AI and big data in governance.
To be clear, I do not suggest that mass killing and destruction are explained by and would not have occurred without these technological developments. These were shaped by multiple dynamics, including but not limited to the Israeli settler movement's aspirations to re-colonize Gaza and expel its residents; the shock, fear, anger and hatred caused by the October 7 massacre of Israeli civilians; Right-wing populist sentiments in Israel predating October 2023, including the dehumanization of Palestinians (Levy, 2025), indifference to Palestinian lives, and rejection of IHL; Israel's engagement in biopolitics, evident in policies limiting food based on calorie calculations, first applied in the 2008 blockade on Gaza; 3 and possibly, erosion in adherence to IHL throughout the war. I do suggest, however, that during the war's first phase, digital technologies played a crucial role in broadening the scope of killing and destruction in ways that are of wider significance for students of AI and big data.
The analysis below relies on both public data and interviews. Public data included public statements of the Israeli army and investigative reports and interviews with officers published in the media. Interviews were conducted with Israeli reserve officers familiar with the use of intelligence AI systems for targeting (including intelligence and Air Force officers), conducted by myself and (separately) by Adam Raz, who generously gave me access to his data. Interviews addressed the definition of targets, the development and introduction of AI systems and their effects on the target production process; and the role played by AI systems within the wider context of the war. In some cases, interviewees refused to address issues they were not allowed to discuss, which I very much respected.
After reviewing the literature on data analytics as a technology of distinction; and on the emergence of targeted killing and the introduction of big data to military targeting, the first findings section offers a solution for the October paradox. It describes the role of big data in the assemblage of what former Israeli Chief of Staff Aviv Kohavi called “industrialized extermination”; how AI accelerated target production; and the circumstances that allowed a large share of Gaza's residents and buildings to be constructed as “targets.” I show how principles developed to legitimate targeted killing of senior commanders in the early 2000s scaled up to destroy entire neighborhoods in full-scale war, highlighting the instrumental role of AI and the automation of intelligence in this expansion. The second findings section discusses various ways AI is used as a legitimation tool. Finally, the third findings section addresses the epistemological-legal challenge of constructing people and places as targets through algorithmic prediction: Intelligence officers did not easily trust AI, and complex trust-building cultural work was required to enable using AI for distinction for indiscrimination.
Literature review
Technologies of distinction
According to the singularization thesis, decision-making and its underlying social sorting (Lyon, 2003) have been delegated to big-data algorithmic systems and now rely on individual-level surveillance (Fourcade and Healy, 2024). This allows for making fine distinctions between individual people and objects, treating each differently, as a unique case (rather than assigning them to preexisting categories). This is evident in the personalization of advertisements, music streaming, social media content, personalized medicine, dynamic pricing in online retail, telematics-based insurance rates, governance, policing, law, social credit, and security (Barry and Charpentier, 2020; Cheney-Lippold, 2017; Einarsson and Ørmen, 2025; Fourcade and Healy, 2024; Kotliar and Grosglik, 2023; Krogh, 2025; Lake, 2017; Latour et al., 2012; Reckwitz, 2020; Ridgway, 2023; Rona-Tas, 2020). In some cases, this results in every individual being treated differently (thus, two social media users are hardly ever exposed to exactly the same content).
Admittedly, this is not always the case: the same technologies used to de-aggregate individuals are also used to reaggregate them (Lury and Day, 2019) into new categories, treating each category differently. As the thriving algorithmic discrimination literature shows, individual scoring may eventually reproduce pre-existing hierarchies and stratification (Airoldi, 2022; Brayne, 2020). Thus, “group-level differences that the law kicks out the door come back in through the window” (Fourcade and Healy, 2024: 243). Yet, algorithms classify individuals into dynamic categories based on constant surveillance of social conduct, not on stable ontological differences. These algorithmically calculated clusters only partially overlap with social groups, even while allegedly representing them (Kotliar, 2020) or aimed at identifying them. They resemble financial derivatives (Arvidsson, 2016): affinities between individuals and clusters are dynamic, meaning both (in)dividuals and the categories to which they are classified are constantly redefined (Lury and Day, 2019; for empirical exploration: Cheney-Lippold, 2017). Moreover, individuals are not simply seen as group members (women or terrorists) but as having a dynamic personalized membership probability (75% women: Cheney-Lippold, 2017). Therefore, big data analytics is often viewed as a technology of distinction. This singularization narrative is also used to make sense of the introduction of these technologies into military intelligence.
Targeting individuals
Technologies of distinction are of interest for armies due to the legal principle of distinction in IHL: the obligation to aim at military targets only and avoid disproportionate harm to civilians. Since 2000, Israel and the USA have openly targeted individual commanders of terrorist and guerrilla organizations through so-called “targeted killing” operations, air strikes on their homes or vehicles, often while surrounded by civilians. While historically, extrajudicial assassinations were limited to rare covert operations against high-ranking leaders, in twenty-first-century Israel, “targeted killing” turned into a declared policy systematically employed on ever larger scales.
This policy reflects three principles. First, the individuation of warfare: fighting between collectives is reframed as policing or prosecution of individuals, who become targets not by their status (membership in an armed force) but by their conduct (Jones, 2020), individual guilt, or the threat they pose (Blum, 2014). Second, the juridification of warfare: due to the growing complexity of IHL, the establishment of the International Criminal Court, and the shift towards asymmetric warfare against terrorist organizations operating within urban environments (that blurs the legal distinction between legitimate military targets and protected civilians), militaries like Israel's and US's have added lawyers to their kill chains (Jones, 2020) to protect their personnel from prosecution. Legal advisors no longer simply concretized the law by setting rules in advance but interpreted these rules in the real-time authorization of targets (determining whether the attack in question is legal according to the principles of necessity, proportionality, and distinction) while interpreting IHL creatively and expansively. Juridification has typically required data collection to establish targeting legality long before the automation of intelligence discussed here. Admittedly, this juridification remained partial: killing without due process, legal defense, or right of appeal, which was often preventive rather than punitive (Guiora, 2013), yet it had consequences. For example, preoccupation with a strike's legality shifted attention from political questions (whether it serves one's long-term interests: Levy, 2023). Third, risk-transfer (Levy, 2017; Shaw, 2002), attempting to minimize the risk to soldiers (as public opinion views their death as an unbearable sacrifice) by transferring it to enemy civilians, e.g., by using aerial bombing, preferably by drones, instead of ground combat or arrests.
Since 2000, the scale of “targeted killing” has grown dramatically, from dozens to thousands, as the definition of legitimate targets was controversially expanded to include non-combatant members of terrorist organizations and, later, civilians who contributed indirectly to the hostilities (Jones, 2020). 4 This expansion created a new challenge that AI systems sought to address through automation: proving the target's connection to the terrorist organization. While this connection was clear when targeting senior commanders, it became ever harder to prove as the range of targets expanded. It was only in the 2023 Gaza War, with the decision to attack tens of thousands of junior Hamas operatives, that the need first arose to examine whether the target considered for assassination had any connection to a terrorist organization, and to which organization. Thus, the use of big data transformed extrajudicial killing targets from singularities (unique, senior individuals known to the targeters) into cases in algorithmically-calculated dynamic clusters.
While “targeted killings” require some intelligence and surveillance technologies, big data analytics was gradually introduced, first to reduce high rates of human misidentification errors (Suchman, 2023) and later, on a wider scale, to streamline killing, reduce its costs and increase its scope, as shown below. While the use of AI to “incriminate” targets relies on the individuation and juridification of warfare, automatic quantification of risks conflicts with fundamental legal principles like reasoning, reflection, and situated decision-making. Yet, it is consistent with big data's hyperindividualist ontology that represents the world as an aggregation of individuated data points to govern atomistic behavior (Lake, 2017).
Critical scholars argue that “targeted killing” relies on an “apparatus of distinction” (Perugini and Gordon, 2017) designed to produce human targets for military purposes by identifying “anomalies” in the relationships between data points, i.e., by monitoring people's behavior and identifying data irregularities (Perugini and Gordon, 2017). Statistical deviations are then marked as normative deviations punishable by death (Huelss, 2019). Even before automation, assassinations were based on identifying life patterns associated with terrorism (Wilcox, 2017). Thus, surveillance and data enable reclassifying individuals from protected “civilians” into enemy targets deemed moral and legal to kill (Bonds, 2019: 442), even at times they do not directly participate in hostilities (Jones, 2020: 182). Bonds termed this “humanitized violence,” a new form of violence that is both a practice (based on surveillance and precision-killing technologies) and a legitimating discourse (which draws on human rights language and portrays its adherents, the US military, as rational, restrained, and humane for struggling to minimize harm to innocent civilians by weighing each strike's predicted collateral damage against its military benefit).
This has a “paradoxical effect” (Smith, 2021): strict adherence to procedures and legal criteria (which prohibit targeting civilians not participating in hostilities and require that harm to civilians be minimal and proportional to the military advantage) and the use of advanced technologies aimed at reducing civilian harm (e.g., facial recognition technologies to prevent misidentification and imaging technologies to estimate collateral damage and calculate proportionality with actuarial precision) increase the legitimacy of and tolerance for killing of civilians, since these calculations render every attack legally justified and civilian deaths regrettable but legitimate. Efforts to minimize “lateral damage” (by warning civilians before attacks and through proportionality calculations) frame incidents with multiple civilian deaths as unintended accidents (Jones, 2020; Levy, 2023).
Digitalization may significantly enhance this legitimation. Algorithmically-calculated scores enjoy an aura of objectivity and trust (Rieder and Simon, 2016) due to the well-documented trust in numbers (the tendency to view quantitative metrics as objective and authoritative: Porter, 1995) and automation bias (the tendency to defer to automated systems, including AI, despite contradictory information: Skitka et al., 1999). By converting uncertainty about potential harm to citizens into measurable “risks,” algorithms may translate moral dilemmas into technical-computational procedures (Smith, 2021) and reduce moral doubts and reflexivity (Suchman, 2023). As shown below, in the Gaza War, AI played a key role in legitimating the upscaling of targeted killings from isolated strikes to mass killing.
To conclude: critical literature shows that “humanitized” violence primarily kills civilians (60%-90% of fatalities are “lateral damage": Perugini and Gordon, 2017), being imprecise (representing success in identifying legitimate targets in terms of “precision” is itself problematic: Suchman, 2020) and narrowing protected categories while legitimating killing. This is highly important criticism of the distinction apparatus. The context of the 2023 Gaza war raises, however, a new question: Do these technologies remain “technologies of distinction” when applied on mass scale, and if so, in which sense?
Findings I: Solving the October paradox
The industrialization of killing: AI as part of an assemblage
King (2024) identifies a gap in the literature that this article seeks to address: While the literature on military AI focuses on dystopian fears of lethal autonomous weapons (“killer robots”), currently these systems are hardly used. While non-AI autonomous weapons are far from new, 5 the main (and understudied) actual military use of AI is intelligence decision support systems. King gives as an example the Ukrainian 2022 attack on Mariivka barracks that killed 600 Russian recruits based on automated monitoring and analysis of social media photos uploaded by Russian soldiers. Yet, the literature (King included) fails to look closely at AI use in intelligence and its implications.
To offer such an account, we must explore the wider context in which AI systems are embedded. Unlike engineers, social scientists do not isolate computer code but analyze it as part of broader assemblages (Gillespie, 2014): “the action, or doing, of algorithms must be understood in situated practices—as part of the heterogeneous sociomaterial assemblages within which they are embedded” (Introna, 2016: 20). Technological deterministic accounts of a given technology's “implications” fail to notice that a technology has hardly any inherent properties: these emerge only from the wider network or assemblage. In our case, the assemblage that shaped AI's implications for warfare includes, inter alia:
International humanitarian law (IHL), legal principles (anchored in legal documents and the history of their interpretation) that set the conditions of legitimacy of military violence. Contemporary militaries (including the Israeli) translate it into formal organizational procedures that regulate target production and approval. One of my interviewees stressed that unlike tank fire, which is not documented and hence grants forces more leeway, airstrikes against intelligence targets are subject to IHL approval procedures. IHL subjects militaries to the principles of distinction (targeting only military targets whose attack provides military advantage, not civilians or civil infrastructure), proportionality (prohibiting attacks that harm civilians disproportionately to the military advantage gained), and precaution. The lethalness doctrine, which turned enemy fatalities from a means to ends such as control of terrain into an end in itself; and enemy fatality numbers into primary measure for evaluating military operations or unit performances, with fighting aimed at killing the enemy fighters to the last (Levy, 2023). This doctrine, emerging in the Israeli military in the early twenty-first century, was officially adopted in 2019, when Kohavi was appointed Chief of Staff. Levy suggests its adoption reflects political changes in Israeli government attitudes: war was no longer viewed as a tool aimed at reaching better diplomatic arrangements but as a way to avoid them. In 2022, Kohavi announced “a central principle, the industrialization of precise extermination.” Combined with IHL and risk transfer, “lethalness” required mass production of legitimate targets for air strikes. Kohavi believed industrializing extermination necessitated the industrializing of target production, this article's topic.
“Targets” is a social construction: while both IHL and internal moral and professional norms oblige armies to aim only at legitimate “targets,” IHL developments and changes in military conventions redefine what is considered a target at a particular time and context. However, targets can be viewed as a not-merely-social construction, one that also involves material objects. 6 This is so, since the procedures used to transform persons or buildings into targets involve various actants joining into a network or assemblage: not only legal definitions, moral concepts, and intelligence officers but also intelligence materials, surveillance technologies producing them, and technologies used to process and analyze them. Since target construction is not-merely-social, technological change may transform it, as shown below.
Ranking as acceleration
How has AI changed the not-merely-social construction of targets? In the Gaza War, Israel used two separate types of AI systems, for infrastructure and human targets respectively. They had different histories but many commonalities.
Interviewees emphasized that both systems did not make decisions instead of humans but accelerate target production in two ways: by fusing information from multiple sources (visual, signal, human, open, cyber, etc.) and making all relevant information about each potential target accessible to analysts; and by ranking potential targets according to their estimated probability. Thus, human analysts can concentrate exclusively on those targets most likely to be approved: “[Assume] you have a billion pieces of information, and (…) your enemy has only 100 targets, you know it, you checked, but you have 10,000 candidates. Now, what do you use AI for? Just for one thing: to sort them by priority. (…) [The computer] took all 10,000 suspects, looked at a few thousand real ones, and [was told to sort out] everything that looks similar in every parameter. Then the machine basically prioritizes, that's all it does. Once it prioritizes, it tells the intelligence agencies: check this and this and this. Assign a work queue. (…) It means the work is being streamlined. But there's no machine that decides” (reserve officer).
This humanistic and humanocentric narrative, also common in IDF statements on AI, downplayed technology's role, reducing it to a tool realizing human goals. IDF described the “Gospel” AI system as merely a “technical tool for the intelligence researcher,” since its traceability and intelligibility allow researchers to examine themselves the intelligence materials on which its recommendations rely. 7 However, even partial delegation of targeting can do more than this: first, meso-level organizational pressures for efficiency and micro-level automation bias may lead human analysts to approve system recommendations almost automatically; Furthermore, tools do not merely realize user goals but also shape them by offering new possible paths of action (Orlikowski, 2007) and encouraging users to use them in certain ways (Davis and Chouinard, 2016). The combination of AI's capacity to accelerate target production, the October 7 shock, and the lethalness doctrine resulted in destruction and killing at unprecedented scale: IDF reported attacking 15,000 targets during the war's first 35 days, with intelligence officers told the goal is “killing as many Hamas operatives as possible” (Abraham, 2023).
Yossi Sariel, former commander of IDF intelligence Unit 8200, argued that acceleration is one of two main contributions of computer automation to intelligence. He claimed that producing more targets is required to exert continuous pressure on the enemy and defeat him, but humans are a bottleneck, as creating so many targets would require thousands of intelligence investigators processing and analyzing data over years. What is required to “blast the bottleneck wide open” is a “human-machine team” that can create a bank of tens of thousands of targets and generate thousands more each combat day (Y.S., 2021).
The second contribution Sariel identified is prediction, defined as “filling in missing information” based on patterns in big data. Prediction is thus not only future-oriented (predicting who will carry out a suicide attack) but mainly present and past-oriented (predicting where rockets were hidden). AI can allegedly use big data to make predictions to be used in target production (Y.S., 2021). This has epistemological significance: it allows “incriminating” people and places based on unknown incriminating information inferred from known non-incriminating information through big data analysis.
Infrastructure targets
The AI production of Infrastructure targets takes place in the Targeting Directorate that was established in 2019, two years after a State Comptroller report indicated that the target bank at the beginning of Israel's 2014 war against Hamas was much smaller than the potential. An interviewee explained the reasons for establishing the Directorate by saying: “you want a bank of quality targets that you can attack and make Hamas surrender,” but in 2014 “It didn't happen, we had to send boots on the ground.” The Israeli military and media described the Directorate and the “Gospel” AI system as the solution. In earlier wars, targets in the bank ran out after a few days or weeks of intensive bombing, leading to a dramatic decrease in fire volume. The Directorate promised to fix this by accelerating and streamlining target production before and during war and thus “transform IDF's destruction capability into an industrial system” that would “destroy thousands of targets every day” (Fishman, 2020). The system fuses billions of data items from various sources (such as intercepted phone calls and aerial photographs), identifies potential targets using machine learning based on their resemblance to previously approved targets, and ranks them for their probability of being legitimate quality targets. These ranked recommendations are then passed to human analysts for decision and to higher officers for approval. Yet, partial automation significantly speeds up target production.
A senior reserve intelligence officer noted: “The Gospel has a very simple user interface that arranges the queue of targets for you according to probability and importance. So the person in the loop simply receives a list set by the machine: how reliable it is to be a target and how important it is. It works through scores. For example, it's a target with an 80% probability, or 30% probability. So the machine recommends. It says: in my opinion, this is a target. The person in the loop takes it, checks the process carried out by the machine, exercises deliberation, because the machine can sometimes make mistakes, and then decides it is a target.” Another officer explained that machine prioritization became necessary due to “exponential growth” in data (following digitalization, datafication and the imperative to collect and analyze all data: Chan et al., 2022; Fourcade and Healy, 2024) that made human analysis unfeasible: “What do we do? Let's bring in 50,000 intelligence personnel? We don't have them. So let's bring in a good computer that will prioritize for us. Basically, that's what the Targeting Directorate was supposed to do.” A jurist involved in target production claimed the system's production rate is 50 times faster than a team of 20 intelligence officers (Brumfiel, 2023).
Interviewees and official army statements stressed that the Targeting Directorate's purpose is “producing large banks” to allow for attacking “thousands of targets in one day,” industrializing extermination. As usual, industrialization was followed by work intensification: as early as 2019, soldiers in the recently-established Directorate reported pressure to accelerate target production by avoiding in-depth examination, with incentives like days off for the most “productive” teams due to a “sense of a shortage of targets.” The short shelf-life of targets was extended by changing procedures, allowing bombing targets months after their production without further examination (Kubovich, 2019a, 2019b).
Accelerating target production is deemed most crucial in wartime: as one interviewee explained, “when a conflict erupts, a lot more intelligence is revealed,” since militants are less careful about information security. Additionally, in wartime, targets may change more quickly than they can be produced. Kohavi noted that AI accelerated target wartime production during a previous conflict in 2021, generating 100 targets per day compared to 50 targets per year produced manually, which was his answer to those who believed “AI doesn't win wars” (Leshem, 2023). This acceleration peaked in the current Gaza war: on the war's 27th day, the IDF announced it has attacked 12,000 targets while simultaneously producing 1200 new targets with its “target factory operating around the clock.”
Human targets: The construction of 37,000 assassination targets
The use of AI to identify human targets has a different story, beginning with AI use for preliminary arrests. In the fall of 2015, Israel faced a wave of spontaneous attacks, primarily by teenagers. Thwarting these attacks was challenging, since perpetrators were not affiliated with terror/guerilla organizational networks. Consequently, the Israeli secret service (Shin Bet) developed a model to predict which teenagers were most likely to carry out attacks by analyzing patterns in social media activity (posts, likes, comments, emojis, new ties) and data from other sources (e.g., location data), assigning every Palestinian teenager a risk score. The model used predictive patterns identified by both machine learning (big data analysis) and human analysts (e.g., new haircuts, as perpetrators often had haircuts shortly before committing suicide attacks). This model resulted in the pre-emptive arrests of hundreds of Palestinian teenagers (Barbing and Glick, 2019; Hirschauge and Shezaf, 2017; Y.S., 2021). Other systems were later developed to identify members of terrorist organizations for arrest and interrogation.
In the recent Gaza War, this strategy was extended to create unprecedentedly long kill lists, deploying predictive policing principles for mass killing while increasing the number of targets dramatically. AI system Lavender assigned almost every Gaza resident a probability score for being a Hamas member, based on factors commonly used in predictive policing (Brayne, 2020) like personal networks (Bob, 2023) and life patterns (e.g., locations and movements) typical of Hamas operatives. 37,000 Palestinians were thus algorithmically identified as probable Hamas operatives (including non-combatants and civilians working for Hamas government) and marked for assassination (Abraham, 2024). Probability scores turn the terrorist/civilian distinction from a binary into a statistical continuum (Weber, 2016). Once a list of probable junior Hamas operatives was prepared, they were tracked by multiple surveillance technologies, and once at home they were bombed together with their relatives and neighbors.
Military sources told Abraham (2024) that intelligence personnel first manually checked the accuracy of a sample of AI-recommended targets but once accuracy reached 90%, two weeks into the war, sweeping use of the system was authorized and individuals it classified as probable Hamas operatives were added to kill lists without further examination (a claim the military denied) apart from checking for sex (as women were assumed to be false positives). Other reports indicate at least some level of human oversight; for example, officers were reported to have corrected AI's misinterpretation of an exam list of high school students as a list of potential militants, which could have led to the misidentification of 1000 teenagers as targets; or a false incriminating machine translation of intercepted phone calls that was identified “by chance” (Biesecker et al., 2025).
Singularity and mass destruction
During the Gaza War, the Israeli government sought as much fire as possible. PM Benjamin Netanyahu reportedly criticized Chief of Staff Herzi Halevi for bombing “only” 1500 targets in the war's first 48 h, rejecting Halevi's explanation that he didn't have 5000 targets by stating “I don't care” (Barnea, 2025). Yet, within the broader institutional-legal context into which AI systems were introduced (the assemblage that included IHL and its target concept), achieving killing and destruction on such a large scale without using AI would have been difficult.
Importantly, The Israeli military relaxed its interpretation of IHL during the war but did not abandon its framework completely, framing air raids as attacks on military targets. CDD (collateral damage degrees) values were raised to 15–20, allowing assassination attacks against each of the 37,000 probable junior Hamas operatives if predicted civilian deaths were below this threshold; CDDs for senior commanders exceeded 100 (Abraham, 2024); “roof-knocking” (warning bombardments) was abolished; and lower-probability targets were considered. Consequently, civilian fatality rates were exceptionally high (Federman, 2024). The attitude toward civilian infrastructure also shifted to maximize destruction: high-rise blocks and university buildings (so-called “power targets”) were attacked for a legitimate military target on one floor but with armament that brought down the entire building (Abraham, 2023). These new standards could redefine most high-rise buildings as legitimate targets. From an IHL perspective, this may well be viewed as breaches of the proportionality and precaution principles.
However, for our purposes, this adherence to the IHL framework may help solve the October paradox: tens of thousands of people and places were categorized as targets, each for different reasons. Each individual person or building was bombed based on singular, specific data associating it with Hamas with a certain probability. As demonstrated below, intelligence analysts sometimes also produced singular stories connecting the data pieces to “incriminate” each target. Yet, multiplied by thousands, this singular targeting resulted in (and was possibly aimed at producing) mass killing and destruction of entire neighborhoods. 8 Admittedly, the high rate of “power targets” (50% of the targets in the first five days: Abraham, 2023) may indicate that military targets were merely used to legitimate and camouflage urbicide, mass killing and destruction of civil infrastructure. However, excuses are sociologically important, as without them some paths of action cannot be justified. We do not know whether pilots in the Israeli Air Force (where the role of lawyers in decision-making is institutionalized for two decades, attacks are meticulously documented, and pilots have long perceived themselves as moral actors based on a strong ethos of adherence to legal procedures, as detailed in the next section) would have obeyed orders to bomb people and places without first constructing each as a “target.” We do know AI systems saved them the dilemma by legitimizing attacks. AI's role was, then, constructing the general as singular, turning almost everything into targets.
Findings II: AI as a legitimation tool
The partial automation of target production provoked internal moral criticism (discussed below), but surprisingly also served as a legitimizing tool. Interviewees stressed it introduced two distinct dimensions of Weberian formal rationality (Brubaker, 1988): adherence to procedures and calculations.
A senior intelligence officer described automation as safeguarding ethical standards by inscribing them in code, transforming them from regulative rules into inviolable generative rules (Schwarz, 2021), thus proscribing their loosening in wartime: “It's an iron rule. You can’t have a target without having two independent sources (…) If you violate it, the machine won’t let you (…) One of the good things about the Gospel is that it is subject to [these rules]. You cannot play with it, like you could say to a human person, ‘Do you hear me? Change it for me’.” By enforcing strict adherence to procedures (here, the norm of multiple sources that widely applies in the Israeli army: Dwoskin, 2024), automation allegedly upholds ethics. Furthermore, he claimed that “When the production of targets is industrialized, the targets produced are more likely to be high-quality” than before, when forces risked “being dragged” into attacking lower-quality targets because “you don't have anything [better] to attack.” In practice, however, only some rules were inscribed in code: the same officer noted the system lets users lower the confidence threshold required for investigation (“If normally a target scored below 80% probability wouldn’t be investigated because it would be a waste of time, [in wartime] you can decide you want to investigate these targets too”). Yet, his narrative mainly presented automation as curbing humans’ tendency to bend rules. Thus, AI takes the use of procedures to legitimate killing of civilians (Smith, 2021) to a new level.
This reflects a broader strategy: beyond the context of AI, interviewees repeatedly made claims to moral status based on adherence to procedures, such as requiring multiple sources of different kinds (“you can't have both be sigint”) and legal advice (“the legal and ethical factor is inherent to the process of target production, like, the people sitting with me in the room are not only target analysts, there's a lawyer, and IHL representatives”). One Air Force interviewee engaged in boundary work, contrasting Air Force attacks where “every bomb is documented” and legally approved in advance with impulsive undocumented tank fire: if a ground commander explains his requests to bomb a house by saying “I just don’t like it. It disturbs the view,” the Air Force will refuse, but “he can ask the tank nearby, ‘throw three shells there.’ A tank doesn’t ask questions (…) Why? Because it's less documented, it's obscured by the fog of war.” This moral frame, in which moral status depends on procedures, not on outcomes, lends itself to legitimize mass killing and destruction.
AI systems were also constructed as ethical through calculations, i.e., performance tests showing they outperform human analysts. One interviewee described how he “compared the performances of AI with a bunch of people, I conducted a real experiment, not historical, live experiment, and we saw that in most cases the AI is more correct than the intelligence officers. That is, if I were to bomb according to what the intelligence officers say, I would bomb more innocent people.”
Another paradox arises: AI's improved reliability may lower misidentification risk per attack, yet, by dramatically accelerating target production and reducing its costs, it enabled a surge in attacks and consequently in accumulated “collateral damage,” increasing the total number of civilians killed. While misidentification became rarer, AI served as a legitimation mechanism for the mass killing of civilians, opening a new chapter in the development of “humanitized violence” (Bonds, 2019).
Another senior intelligence officer described big data's legitimating role as both tactical and strategic. Tactically, it legitimated the flattening of urban areas before ground maneuver: “This is the humanitarian maneuver method, which is to some extent more ethical towards our soldiers. (…) When [they] reached certain areas, they had already been completely flattened.”
Constructing buildings in maneuver areas as targets, he claimed, “is one explanation for the unusual scope of targets attacked.” But strategically, AI morally and legally legitimated the wider project of mass killing and destruction: “Let's say the system gave me a target, a building with a Yakhont missile underneath it. And I know that if I bring down the complex with its residents, hundreds of civilians will die. So the target actually provides me with the excuse to carry out extermination, right? It gives me the moral cover. A rich target bank is effective. But if you zoom out, you must ask yourself: how does it help me? (…) We really like to speak the humanitarian maneuver language. We maneuver, meaning we only kill the bad guys. The targets really help us market it that way. But eventually, does it really matter whether there were targets underneath all the houses we brought down in Gaza or not? Now, at this moment of the war, when Beit Hanoun is a pile of dust, what does it actually matter? As mentioned, it's important for moral justification.”
Both denying AI agency (portraying it as a tool that merely streamlines human decision-making) and stressing it (portraying it as a strict rule enforcer or superhuman analyst) were used then for legitimation.
Findings III: Trust in numbers as a legal-epistemological challenge
Why weren't AI recommendations “sent to the airplanes"?
An interviewee told me that senior officers “wanted to go another step further” and let AI “actually produce targets. We haven't reached this point, I think it's still far-fetched.” A senior officer had even proposed to automatically strike objects algorithmically identified as targets with the highest probability, “connecting it to the airplanes”, as a deviation of “less than 1%” constitutes a “legitimate error rate, so let's send them directly, automate it completely.” My interviewee insisted this proposal was rejected but believed it will eventually happen, “once the cost of failure will be too high,” although he stressed this means carrying out a known percentage of attacks on illegitimate targets that constitute “war crimes.” However, Abraham (2024) claimed that in the Gaza war, humans on the loop were reduced to “rubber stamps,” approving targets in as little as 20 seconds, without examining raw data.
The decision to avoid automating decision-making had several reasons. First, an unease about being those setting a precedent, out of “concern for the international consequences of… a machine that decides to attack (…) no one did it before” (senior reserve officer). Second, the need for accountability: IHL requires human(s) responsible for each decision; “Since it's a matter of life and death, we intentionally kept a human in the loop, so we won't have mistakes without owners and end up finding the machine accountable. We did this although it's theoretically possible to connect the target generation system directly to the strike system. For instance, any target scoring above 90% could be automatically sent to the Air Force for attack.” Third, moral defense. Even if humans approve targets within seconds and rarely overrule AI, a human in the loop makes decisions seem like products of human moral deliberation. One officer stated: “The classic question I'm asked is whether AI replaces people or whether it causes irresponsible actions. The answer to both questions is ‘absolutely not’. (…) The hand on the trigger is literally a human hand.” Like other interviewees, his account reduced AI's actancy to organizing the queue, while stressing “meticulous procedures": “There's no crazy machine here that we let run wild.” Moreover, due to legal and ethical concerns about delegating life-and-death decisions to algorithms, the tolerance for false positives is lower than with human decisions, setting a standard much higher than AI's actual performance: “Even a 99.1% result is not good enough for the military. Morally, in terms of the error threshold it tolerates. No machine achieves it, machines yield much lousier numbers.” Hence, he stressed, “the military is not prepared, neither morally nor practically, for any machine to make decisions.”
Storytelling: Opening the blackbox
Finally, another factor preventing complete automation was the epistemological and moral views of intelligence officers and the tension between AI's statistical knowledge and what they were trained to perceive as valid knowledge required for target production, as the construction of targets is not merely social, but neither is it merely technological.
Intelligence officers believed that for targeting decisions to be professional, moral and legal, they must be capable of explaining how they were reached and challenge them through counter-interpretations. This perspective posed a challenge for the technical teams developing big data analytics systems: as one interviewee noted, failure to open the blackbox and convey a human story often led to system discontinuation: “But the user won't trust the probabilities the model throws out, 80%, 90%” Interviewer: But they are supposed to organize his workflow, right? starting with… “Listen, it's a very long process until he trusts these numbers and this workflow… how does he trust it? Only once he understands the ranking, what it's based on, and once he's tried it over time, seeing it's consistent with the story he has in mind. If it's inconsistent with this story, he'll be the first to abandon it, and that's it. The number of systems that were created and abandoned, during the war and before it, is enormous. Because the user didn't trust it, he said ‘I don't believe it’.” Interviewer: Systems abandoned after a period of use or that weren’t even put to use? Or both? “Both.”
Epistemological norms prevailing in the field act as a counterforce restraining automation bias and trust in numbers, in line with findings on radiologists’ trust in AI: when faced with contradictions between their own evaluation and AI's, radiologists tended to disregard AI whenever they could not understand the reasoning behind its differing conclusion (Lebovitz et al., 2022). Another interviewee explained that approving a target necessitates “a proof,” which must be “a human story” that can be challenged with counter-interpretations: “We have IHL guys with us every time we decide on a target: ‘tell me, how do you explain this tomorrow in The Hague?’ (…) Every researcher must defend his thesis before someone more senior and must give explanations, that person says to him, ‘Tell me, but maybe it's like that?’, he says ‘No, I checked it.’ And sometimes he doesn't know everything and says ‘Listen, to me it makes enough sense to be a target.’ And sometimes he's negligent and says, ‘Well, I don't have the energy, put it in the pool for me.’ All these things are human.”
A senior reserve officer explained that while AI systems can identify and connect details, they struggle with the storytelling needed to grasp the overall meaning. For instance, they can recognize that a picture contains “an adult man, a small child, a car, and a football,” but have difficulty determining “what's happening in the picture? Semantically, what's the story? The story could be that the child kicked the ball into the road and the man ran to save him, or that the child is playing while the father picks him up from kindergarten. (…) based on many clues, any person will tell you whether it's this or that.” In intelligence contexts, he noted, an AI system could identify a location visited by multiple militants and rank it high, failing to consider it could be “a very good restaurant where militants dine,” or that a captured conversation about terrorism might be from a movie shown on television.
Another interviewee, who served in a technical role, discussed how agency is distributed between AI and intelligence analysts: “usually the intelligence people know something, and you try to find that something,” but in other cases the systems are used “to find correlations (…) between all kinds of variables that until now you didn't see as interesting (…) Obviously I show the intelligence people a lot of data and a lot of correlations and tell them, ‘Look, this is interesting’.” Then the intelligence analysts try to open the blackbox and produce a narrative that explains what the model does and why it works, which is necessary for the model to be trusted and put to use. “I show them the very specific data points, [variables that predict something].” However, he stressed that in these instances, system's predictions are not used to produce targets before intelligence analysts opened the blackbox and provided an explanation for why these specific variables included in the model predict outcomes, contrary to the radical empiricist-epistemological promise of AI prediction to render explanation redundant (the so-called “end of theory": Kitchin, 2014). The interviewee suggested that human narratives are still required “when AI is integrated into mission-critical systems.” Generally, the challenge in machine learning is that you can predict, you can tell something very well about your tested, about your evaluation set, but you can't tell what will happen in reality. It's speculation based on all kinds of statistical assumptions (…) and in places where the cost of error is very high, and here we're talking about very-very high costs of error, harm to human life, damage to [the army's] image, [so] you don't just say ‘Oh, it seems to work in my tested, let's put it in production and now it will produce targets (…) that we'll bomb, because you want to understand what's behind this thing, like, what's in there. Because with all you trusting your data-science people, for that matter, and having collected the tested well, you can't tell how good it is (…) you must explain at the most precise level possible what the variables say, and why the model predicts what it predicts.”
Building trust, camouflaging AI
Trust in numbers is not as omnipresent and automatic as the literature may lead us to believe. In our case, big data systems’ end users (intelligence analysts) were initially suspicious of automation and algorithmically-calculated probabilities. A developer I interviewed criticized them for distrusting AI systems due to their “fear of error” and excessive caution (“They need a level of certainty that is not 100% but 200%, and there's no such reality”). To overcome these suspicions and generate “trust in numbers,” system developers engaged in cultural work and devised a sophisticated strategy: disguising the role of AI and emphasizing human involvement. “Intelligence people don't use AI, they use tools built for them, and from their perspective (…) if there's AI there, in my experience, they don't trust it. In fact, our way to get them to use it was not to tell them ‘it's pure AI,’ but rather to make sure that the [intelligence person] who goes over a target must see that Sharon went over it earlier, they can see the actual name of the person who looked at it, and they need to know Sharon. They need to know that Sharon is the best in targets and therefore they can trust it”
In this example, they leveraged the trust intelligence personnel had in a highly-appreciated former intelligence officer (“Sharon”), who was promoted and transferred to the technical team, where she labels data and validates algorithmically-generated targets before they are passed on to intelligence and legal personnel for approval. Personal trust in Sharon can tip the scales: “We saw that when there's a Sharon, they trust it, and when they're told ‘it's just the model,’ they don't [and start the process manually from scratch].”
Conclusion
By looking closely at the use of AI in the Gaza War, this article offers several wider contributions for the study of AI. The first contribution is complicating our view of AI as a technology of distinction. Sociologists and anthropologists have known for long that algorithms are inherently cultural (e.g., Seaver, 2017), yet, being technologies of distinction is viewed as inherent to big data analytics and AI, regardless of cultural context. The Gaza War case complicates the picture, showing that while distinction capabilities are core features of these technologies, they may produce opposite social effects, depending on the wider assemblages of which these systems are part. In our case, it produced distinction for indiscrimination: AI was needed not to personalize treatment but to justify uniform treatment (bombing) by creating personalized justifications for the targeting of nearly every building or probable Hamas operative and to accelerate incrimination and target production to unprecedented levels without completely departing from the IHL framework, legitimating dreadful mass killing and destruction by relying on allegedly objective probability calculations and procedures inscribed in code.
A second contribution is demonstrating why critical studies of the social impacts of algorithms must go beyond criticizing their errors and biases. The latter have been the main foci of critique and for a reason: despite their claim of neutrality, AI systems are prone to errors in general and to reproducing inequality due to bias in particular, as their training data documents and reflects human biases (e.g., Brayne, 2020; Noble, 2018; for critique: Zajko, 2021). Specifically, errors and biases are main concerns in debates on AI in warfare, including in the Gaza War (e.g., Australia, Canada, Estonia et al., 2024; Sylvia, 2025). However, my analysis shows that even a perfect, flawless and bias-free AI system could lead to the same horrific results simply through dramatic acceleration and cost reduction of target production: attacking tens of thousands of intelligence targets, which was practically impossible to produce before AI, leads almost inevitably to mass destruction and mass killing of civilians, and may be used to legitimate them, as each attack may be allegedly justified in legal terms, regardless of their dreadful combined effect as a whole. Since delegation of decisions to AI may have devastating social impacts even when models work perfectly well, effective social critique of AI cannot afford to focus solely on bias and errors.
The mass killing of Palestinians and the destruction of Gazan cities were politically-motivated and not solely caused by AI, yet AI was necessary to enable thousands of air raids within the IHL framework and hence within the norms that pilots considered professional and moral, providing internal legitimation to prevent possible resistance. This legitimation potential may be used beyond the Israeli context, as similar systems are being adopted by other militaries officially committed to IHL. Digital technologies enabled Israeli intelligence to generate targets perceived as legitimate at unprecedented speed, turning tens of thousands of individuals and buildings into targets due to changes in the not-merely-social construction of military targets. Acceleration may sound like a quantitative change, but it brought about a dramatic, qualitative transformation.
AI's role in legitimating this mass killing reminds us that features and affordances do not characterize technologies in isolation; they are informed by the legal, structural, cultural, and moral contexts and by the heterogeneous networks or apparatuses within which technologies operate. Addressing this context is crucial to understand how it could produce horrifying effects, but also, why it was impossible to eliminate the human in the loop, and why the cultural work of meaning-making by humans was necessary to legitimize the partial automation that led to this unprecedented acceleration of target production.
Footnotes
Acknowledgements
I wish to thank Adam Raz for generously granting me access to his data and Yagil Levi and the Big Data & Society's anonymous reviewers for their helpful comments and suggestions.
Ethical considerations
The study was approved by the institutional Ethics Committee on February 26, 2024.
Participants were given all relevant information and gave their informed consent verbally, to ensure their anonymity remains protected due to the topic's sensitivity.
Funding
The author received no financial support for the research, authorship, and/or publication of this article.
Declaration of conflicting interests
The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.