In April 2024 international news sources circulated a video of a talk delivered by a senior Israeli military official during a tech conference at Tel Aviv University about a year earlier. The speaker, whose face is not shown, introduces himself only as “Colonel Yoav,” the head of data science and AI for Unit 8200, an elite outfit within Israel’s intelligence directorate. He is there to talk about the force’s embrace of artificial intelligence for the purposes of, as one slide in his presentation puts it, “finding new terrorists.”

During Israel’s May 2021 assault on Gaza, Colonel Yoav explains, his unit had used AI to generate two hundred targets, a process that in the past would have taken “almost a year.” In comparison, he suggests, the algorithm could identify likely members of militant organizations in practically no time: “Let’s say we have some terrorists that form a group, and we know only some of them. By practicing our data science magic powder we are able to find the rest.”

When Israel again began bombing Gaza after the Hamas-led attacks of October 7, 2023—commencing a war on the Strip that leading experts and human rights organizations have termed a genocide—it used AI to increase its surveillance and killing capacity. In lengthyexposés published in the Israeli/Palestinian outlet +972 Magazine, the journalist Yuval Abraham identified two separate machine-learning algorithms that were churning out lists of people or places to bomb, enabling the Israeli air force to strike targets at an unprecedented speed and scale. The first, called Habsora, or “The Gospel,” offered up locations that were supposedly linked to militant groups or their members, though sources told Abraham that the majority of its suggestions were private residences. The second, called Lavender, made “kill lists” of individual people, assigning scores that approximated how likely they were to belong to a militant group. Intelligence officers told Abraham that, at the start of the war, the military established that operatives could subject Lavender’s recommendations to only the barest scrutiny, even though it had been determined that the program made errors in around 10 percent of cases, sometimes suggesting targets with no Hamas affiliation. “I would invest twenty seconds for each target at this stage, and do dozens of them every day,” one officer said. “I had zero added value as a human, apart from being a stamp of approval.”

The use of these systems, combined with shifts in policy and tactics—like increasing the number of noncombatant fatalities permitted as collateral damage in air strikes, attacking at night when those marked for assassination were at home with their families, and categorizing money changers and lookouts as valid targets—contributed to the bombing’s astronomical toll. At the height of the air campaign, according to the Israeli historian Adam Raz, the Israeli air force was striking Gaza as frequently as twice every minute. “The political leaders were interested in massive destruction,” one intelligence reservist who served after October 7 told me in July 2025. “If you have the ability to produce a massive scale of targets rapidly…even if the intelligence and the air force did everything professionally, it will still cause enormous amounts of damage.”

The Gospel and Lavender are the results of more than a decade’s worth of Israeli military experiments in big data and machine learning aimed at tracking, arresting, or killing Palestinians. One of the Israeli military’s first forays into the subset of technologies known as artificial intelligence was a project dedicated to predictive analytics, which pulls patterns out of data to make suggestions about the future. The system—a collection of algorithms that the military refined over time—claimed to be able to identify potential “lone wolf” assailants before they carried out terror attacks in Israeli cities or settlements. By the mid-2010s such incidents had become increasingly common, and military officials feared they were facing down another intifada.

The technology they developed in response was a blunt instrument compared to the large language models and neural networks that predominate today. Rather than synthesizing troves of information—geolocation tags, photographs from satellites and CCTV cameras, phone calls, text messages, and internet browsing activity—and returning tidy recommendations, the lone wolf algorithm relied on narrow datasets drawn from the telecommunications records and social media accounts of known assailants. Developers treated certain features of their profiles (such as, for example, their marital status) as predictors for carrying out an attack and trained the algorithm to look for similar indicators as it sifted through data from Palestinians in the occupied territories.

The algorithm would assign a score between one and ten, which supposedly corresponded to the threat a person posed to Israeli national security. In effect, it divided Palestinian society into percentiles: “You basically wanted to pick the top 10,000 out of the entire population,” an intelligence officer who trained conscripts in the late 2010s recalled to me. Those with higher scores were marked out for further surveillance or arrested outright, allowing the military to rationalize monitoring or detaining civilians with no prior security record. Under this logic, the officer said, “every Palestinian is a suspect.”

In the years after it deployed that surveillance system, Unit 8200 devoted some of its subunits to AI development. (The Washington Post has also reported at length on the transformation of the unit from a small wiretapping division to one of the largest digital surveillance outfits in the world.) Soldiers who had been trained for analog intelligence work found themselves tasked with the kind of menial labor that tech conglomerates usually outsource to low-wage workers overseas: they spent hours tagging phrases to render transcripts and messages searchable by keyword, correcting automated translations of communications, and engineering user-friendly databases to make the information accessible to combat troops. Eventually this data came to feed the platforms that have expedited target generation not only across the West Bank but also in Gaza and southern Lebanon. In 2023 the director general of Israel’s Ministry of Defense, Eyal Zamir, announced an intention to turn Israel into an “AI superpower.”

It is difficult to measure the material effects of these efforts—both because they are shrouded in secrecy and because the cluster of technologies that bears the name “artificial intelligence” has been the subject of so much hype. But by now it is clear that in addition to its centrality in the assault on Gaza, this technology is also being deployed in other conflicts: Israel and the US used a host of AI-assisted surveillance and targeting systems to intensify their attack on Iran, relying on the programs to churn out a steady stream of targets for bomber pilots and cruise missiles.

Far less remarked upon is the fact that Israel’s expanding algorithmic arsenal has also changed the day-to-day administration of the occupation across the Palestinian territories. Today Israeli soldiers and border police are armed with tablets and smartphone cameras as well as assault rifles as they patrol East Jerusalem and the West Bank. They rely on biometric databases, keyword queries of intercepted communications, and programs that scrape social media to guide arrests. These tools allow soldiers to access the intimate details of Palestinians’ private lives, like who someone called and what they said, where they went and at what time, how they used the internet or social media, and what they photographed on a given day. This surveillance dragnet gives the military new rationales for longstanding practices of mass policing and detention, which officials tout as precisely targeted even when their effects are indiscriminate and brutal. “We always had surveillance, we always had mass detention,” Sahar Francis, a Palestinian lawyer and longtime human rights defender, told me in mid-December. “Now it’s just much easier for them to arrest more people.”

Since Israel outfitted combat troops with AI-assisted tools in the early 2020s, arrests of Palestinians in the territories have steadily risen, and since October 7, 2023, they have increased exponentially. The number of Palestinians in administrative detention, a legal category that allows Israeli authorities to hold anyone deemed “a threat to national security” indefinitely without charging them with any crime, almost doubled between 2020 and 2022, from 355 to 791. By the end of 2025, 3,329 Palestinians were being held without trial—an all-time high. The introduction of new technology has coincided with the rise of ultranationalist political forces: Itamar Ben-Gvir, Israel’s minister of national security and a member of the Jewish supremacist Otzma Yehudit party, controls the Israeli police, national guard, and border police, the latter of which operates in East Jerusalem and the West Bank; he has advocated for rewarding Israeli settlers who shoot Palestinians with medals and subjecting Palestinians convicted of carrying out lethal attacks against Jewish Israelis to the death penalty.

Over the last few years I have spoken with dozens of Israeli reservists and veterans who were involved in the army’s automation efforts. Their testimonies, published here for the first time, offer a newly detailed view into how the AI targeting platforms that the Israeli army has used since October 2023 emerged from ten years of algorithmic surveillance experiments across the occupied Palestinian territories. Of the eight who agreed to be anonymously quoted in this piece, some remain in the military as career officers, while others have become start-up founders or academics. To understand how Israel integrated AI into the way it wages war, they suggested, one first has to see how it came to use these tools to control every aspect of Palestinian life. “The lone wolf system was really the predecessor for the systems we’ve seen in Gaza,” Sebastian Ben Daniel, a computer scientist who publicly revealed Israeli officials’ use of predictive algorithms to monitor Palestinians in a 2015 exposé, told me. “It’s how the military got the money, the support to keep developing them.” At first they saw the technology as a tool to arrest people, he said, but “the next step is to use it to kill people.”   

The Israeli military’s turn toward big data was inspired by the United States, which built one of the largest surveillance dragnets in the world after September 11, 2001. The National Security Agency began vacuuming up Americans’ internet activity, emails, and calls—what the agency’s former head Keith Alexander called the “whole haystack”—and storing them in an enormous data center in the Utah desert. Edward Snowden’s 2013 leak detailing these extralegal surveillance programs sparked a national reckoning with the limits of constitutional privacy protections; the information the government had mined exceeded even what was permitted by controversial provisions within the PATRIOT Act. But by then the pile of data had already seeded predictive analytics as a new and lucrative industry. Companies like Palantir and Dataminr promised fantasticcapabilities, suggesting that their software could pinpoint terror attacks in western Europe before they were reported or track down insurgents in Afghanistan planting roadside bombs. They peddled their wares to governments, police departments, immigration authorities, and militaries around the world, which increasingly embraced mass surveillance.

There was little verifiable evidence that these technologies were accurate, however, and the reasons for skepticism were manifold. Data scientists have long warned that the statistical models that scaffold these systems are, at best, flawed. They can “make good predictions if nothing else changes,” Aravind Narayanan and Sayash Kapoor note in their book AI Snake Oil. But the moment these machine learning systems encounter a reality that differs from the data on which they were trained, they’re liable to make mistakes—and the world, after all, is highly variable, as is the information it produces.

An algorithm might, for example, be taught to identify members of a foreign military based on location and telecommunications data; but the people who travel in and out of military bases and place phone calls to army commanders could include civilian journalists and humanitarian workers. Or it might learn to pick out signs of radicalization in text messages and end up implicating anyone who uses common words like “resistance” or “martyr.” As the anthropologist Lucy Suchman observes, most contemporary counterinsurgency operations feature “incredibly complex relationships between the so-called insurgents and civilian populations” that elude the rigid logics of machine learning—which some journalists and researchers argue is a large part how US intelligence agencies came to place thousands of noncombatants on kill lists in Pakistan, Iraq, and Afghanistan.

None of this deterred Israel’s intelligence agencies, which by the mid-2010s were eager to adopt predictive systems to pinpoint targets for surveillance, detention, or assassination. This was at the height of a wave of lethal attacks carried out by young Palestinians from the West Bank and East Jerusalem, whom security officials called “lone wolf” assailants because they acted alone and typically had no prior record of militancy. They had come of age as hope for a permanent peace plan disintegrated, and their futures felt increasingly foreclosed by dwindling job prospects, crackdowns on travel, and wider economic desperation in the occupied territories. They found their way to Jewish neighborhoods in Israeli cities with kitchen knives stashed in backpacks or pockets, or rammed cars into bus stops in East Jerusalem or outside West Bank settlements. At the height of the violence an average of 1.8 such attacks occurred each day.

The man who oversaw the creation of the system intended to prevent future attacks was Yossi Sariel, a brigadier general who enlisted in military intelligence in 1997 and would come to command unit 8200 in 2021. In the early 2000s he was an enthusiastic member of “the Choir,” a group of spies who advocated for Israeli intelligence to embrace Silicon Valley’s machine learning tech. According to an ex-security official who served above him, Sariel made a name for himself in the early 2010s when, as the commander of 8200’s operations in the West Bank, he headed the effort to create an expansive facial recognition database nicknamed “Google Ayosh.” (“Google” referenced the ease with which someone could be identified by plugging in a picture of their face; “Ayosh” is the Hebrew acronym for Judea and Samaria, the biblical name the Israeli government uses to refer to the West Bank.) It was the lone wolf algorithm, however, that would cement his reputation as one of the most radical tech-accelerationists within Israel’s security agencies.

Israeli intelligence units were uniquely positioned to develop such a system in the Palestinian territories: there are virtually no technical or legal limitations on the military’s ability to mine Palestinians’ calls, texts, and browsing history. Palestinian telecommunications providers are forced under the terms of the Oslo Accords to rout all calls and messages through Israel, and the majority of Palestinian Internet access is also dispatched through servers within Israeli territory. Meanwhile, military rule denies Palestinian civilians in the territories the privacy protections nominally afforded to Israeli citizens.^* In response to the lone wolf attacks, the Israeli army began systematically collecting the enormous amount of data at its fingertips—an effort that grew with each passing year. “It was like a western military’s dream,” one intelligence officer recalled when we spoke in January 2025.

At first, Israeli intelligence units didn’t have the capacity to sift through the data piling up on military servers. “There was always this feeling that we had so much more information than we knew what to do with,” one person who served in 8200 between 2014 and 2016 told me. He recalled high-ups’ enthusiasm for the data analytics firm Palantir. “The kinds of things Palantir was doing, we wanted to do,” he said. “It was a very experimental space.” Though this person didn’t work on what became lone wolf, he spent time building machine learning programs that could propose marks for the air force to bomb in southern Lebanon, seemingly a precursor to the Gospel targeting system that has since been used in Gaza.

The lone wolf system, which by 2016 was being used across the West Bank, was the army’s answer to this information overload. It sorted through the data and identified potential attackers using criteria such as whether someone made a high number of phone calls in the middle of the night, mourned a family member killed by Israeli soldiers on Facebook, or messaged a friend that they wanted to kill themselves. Those with the highest scores were subject to closer surveillance. Individuals who seemed to pose an immediate threat to Israeli national security might be dragged in for questioning (by Israeli forces if they lived in an area directly controlled by the military, or by the Palestinian Authority’s security services—acting under the orders of Israel’s Shin Bet—if they lived in one of the West Bank cities technically subject to PA jurisdiction).

Indiscriminate policing and mass incarceration are nothing new in the Palestinian territories: before the arrival of AI, the Israeli military could and did, for example, respond to stone throwers by arresting every Palestinian man between the ages of seventeen and fifty in a given village. Displays of martial power were, in the words of soldiers who gave testimonies to the Israeli NGO Breaking the Silence, meant to make the army’s “presence felt,” in effect reminding Palestinians they were living under occupation. But the rise of machine learning allowed the army to paint these shows of force with a veneer of technical legitimacy.

The lone wolf system made it possible to sift through reams of information at superhuman speed; Sariel would later boast that using machine learning to pinpoint “lone wolves” allowed intelligence agencies to operate at a pace that would have taken “20,000 analysts 20,000,000 years.” (He did not divulge the scientific method through which he arrived at these numbers.) This, in turn, gave the army an incentive to collect even more data. In the late 2010s, one 8200 veteran told me, the military began automatically recording calls made by Palestinians who had been flagged by algorithms for further surveillance. Israeli authorities also stepped up other forms of both digital and physical surveillance, mounting CCTV cameras and license-plate scanners over the roads crisscrossing the West Bank and East Jerusalem.

According to sources who either helped engineer the system or used it to surveil Palestinians, Israel expanded its use of semi-targeted “trojan horse” hacks in these years. Intelligence units would use phishing attacks to propagate messages with links to malicious software; if clicked on, the links would infect smartphones with programs that took over a device’s operating system and raided data from encrypted messaging applications, like WhatsApp or Signal. More advanced versions of such software can even turn on a phone’s camera and microphone to covertly record.

These attacks would plant “cheap spyware…on hundreds or even thousands of devices,” one person who served in 8200 in the late 2010s recalled. They could be based on “areas or based on lists of semi-suspected people,” and gave the army “access to the WhatsApp communication of thousands of people and by proxy also to the WhatsApp communication of everyone talking to them.” (Another source who was part of the unit at the same time precisely echoed this account.) Reservists told me that the military has continued to carry out these hacks in the West Bank and Gaza Strip on a regular basis ever since.

At the time, military spokespeople framed all these efforts in humanitarian terms: “Unlike members of Hamas or Islamic Jihad, if you arrive at the child’s house a week before the attack, he still doesn’t know he’s a terrorist,” one officer said at a press conference on a West Bank military base in 2016. But the ex-security official who served above Sariel at the time said that in practice the lone wolf system worked to criminalize an understandably widespread sense of despair. “Most of the people [we arrested] were completely innocent,” he said. “It turned out they were suicidal for non-political reasons; they were just fed up.”

In April 2017 the army and security services declared the predictive policing program a success. Officials told Haaretz that the Israeli security forces had thwarted 2,200 attacks and arrested four hundred Palestinian soon-to-be assailants. Another four hundred, they added, had been arrested by the Palestinian Authority on Israel’s behalf. Some questioned whether these numbers were cause for celebration. “How is it possible for Israeli security forces to arrest young Palestinians over things they have not done? How can they be convicted of a crime when they haven’t done anything?” Ben Daniel, the computer scientists who first reported on the program, asked in a Haaretz column that year. “The repeated claim that use of the system has achieved substantial results is an amazing one that is impossible to verify.”

In 2019 Sariel took a sabbatical at the National Defense University in Washington, D.C. His stint there coincided with the beginning of the AI boom. That year Microsoft invested $1 billion in a start-up called OpenAI; other technology giants were also pouring enormous amounts of cash into experiments with large language models. One professor who got to know Sariel during this period later recalled his enthusiasm for AI to The Washington Post: “Yossi was in this world of, ‘This thing is moving fast, faster than anybody realizes. And we better get everybody on board.’”

Sariel spent part of his sabbatical writing a manifesto, which he would go on to self-publish in 2021, titled “The Human-Machine Team: How to Create Synergy between Human & Artificial Intelligence that will Revolutionize Our World.” The text combines clunky citations of the Talmud with TED-talk-style distillations of military history. In its vision, a “merger of human intelligence and artificial intelligence” renders security agencies capable of predicting any potential threat. It urges militaries to prepare for a near future in which the abilities of artificial intelligence will surpass those of human beings; Sariel recommends that armies replace up to 80 percent of their intelligence analysts with algorithms, which he promises will dramatically expand their ability to surveil, identify, and kill targets.

In sketching his vision, Sariel—whose writing tends toward bravado rather than the discretion one might expect of a spy chief—also reveals details about a number of then-classified Israeli military technologies. He suggests that “human-machine teaming” in national security agencies could yield a “smart border” that relies on facial recognition; Israel rolled out such a system across the West Bank in the late 2010s. He talks up the potential of algorithms to “catch lone-wolf terrorists” based on “prediction” with “big data.” And he boosts the idea of a “targets machine” capable of sifting through information about where militants operate and suggesting “thousands of new targets every day.” In this sense, “The Human-Machine Team” promises what most proponents of AI promise: that algorithms will carry out complex tasks more efficiently and precisely than humans. When it comes to marking out people for assassination, Sariel assures readers that machine learning will “find the right targets at the right time.”

The head of Unit 8200 would not typically publish a treatise on the intelligence corps’ work; the very identity of that person is supposed to be a state secret. The text was authored anonymously, but, in what would become a much-discussed lapse, Sariel included an email address on the copyright page that linked to his full name, personal Google account, and calendar. (I first heard about the book when I turned up at a military intelligence archive in 2021 asking about 8200’s embrace of big data and machine learning. “It’s in English and you can download it on Amazon!” a librarian suggested cheerily.) In early 2024, The Guardian journalists Bethan McKernan and Harry Davies discovered the text online and quickly unmasked Sariel as the commander of 8200.

By then he had led the outfit for three years, putting his theories into practice. As ChatGPT became arguably the fastest-adopted technology in consumer history, he embraced the prevailing wisdom of the industry—attempting to create more and better AI capabilities by building larger large language models, trained on bigger sets of text-based data. For Israeli intelligence units, obtaining bigger data sets meant even more dramatically expanding surveillance of the Palestinian territories. The year Sariel took over 8200, the army began storing virtually every call and text message sent from the West Bank and Gaza on Microsoft cloud servers for at least a week. By 2023, according to one reservist, data of interest—like calls placed by someone Lavender had marked as a militant—were being kept for much longer.

Using automated translation and transcription tools, 8200 operatives turned these private communications into searchable text and made the resulting databases easily accessible to anyone linked up to the army’s cloud. The amount of information at the average soldier’s fingertips was stunning; Abraham, of +972, has reported that under Sariel the IDF began storing up to 200 million hours of audio alone. At the same time, Sariel established subunits devoted to engineering more advanced machine learning algorithms to sort through all the new data—including Lavender and the Gospel.

A source who served as an 8200 officer in the late 2010s and early 2020s remembered how an American strain of techno-optimism suffused the unit’s culture. Sariel would echo talking points about transhumanism or a coming human-machine singularity from Silicon Valley CEOs, he said, or direct soldiers to watch videos featuring Elon Musk and other American tech executives between workshops on natural language processing. But it wasn’t just Sariel: the former officer recalled the general “excitement mostly from the top” about AI’s disruptive potential. Commanders “would talk about how in three years…everything would be automated, they wouldn’t need translators, people with language skills; all [intelligence] would just be based on AI,” he said.

Sariel capitalized on that excitement. He found a valuable ally in Aviv Kochavi, chief of the general staff for the military, who had pledged to make the army more “efficient, lethal, and innovative” upon assuming his post in 2019, and had called the IDF the “biggest start-up in Israel and perhaps the entire world.” With Kochavi’s blessing, Sariel pumped resources into scaling up 8200’s products. One unit officer who worked on prototypes of generative AI systems recalled going on a “road show” to assess the technological capacities of various Israeli military bases. Another unit veteran said engineers won internal awards for making interfaces more user-friendly. One interface, called Flow, allowed soldiers to search across surveillance databases by name, phone number, or location, and send themselves alerts whenever a person of interest placed calls or sent a message. 

Sariel also made it his mission to encourage other units to use the platforms 8200 was developing. Ground troops patrolling the West Bank were given smartphones and began stopping civilians on the street and scanning their faces to determine if they should be detained based on an automatically generated security rating. New, specialized units equipped with drones and targeting platforms were dispatched to southern Lebanon to identify and kill Hezbollah operatives. Air force officers began relying on algorithmically generated lists of targets to ramp up drone strikes across the Gaza Strip—The Jerusalem Post dubbed Israel’s May 2021 assault on Gaza “the world’s first AI war.” In those years, the number of people killed by IDF fire skyrocketed: 323 Palestinians were killed by Israeli security forces in 2021, the deadliest year in the occupied Palestinian territories since Operation Protective Edge, a six-week Israeli offensive that pummeled Gaza in 2014.

Most of the reservists I have interviewed believe Sariel’s work helped thwart terrorism by making intelligence units more precise and the army as a whole more powerful. As one 8200 reservist put it: “Given all the information that the army has, it’s necessary to use this stuff as a tool…and now we can do so in a way that can be accurate and lethal, and efficient.” One senior officer who led an engineering unit within 8200 throughout the war on Gaza emphasized what he sees as the Israeli military’s unique mandate to innovate. “The big tech companies don’t have the data we have, they don’t have the material we’re sitting on,” he told me. “We have the biggest amount of spoken Arabic probably in the world.”

Others expressed anxiety about how the technology was used across the army, especially since the Israeli military’s embrace of automation dovetailed with the country’s social and political shift to the extreme right. Through the 2010s, as the Israeli sociologist Yagil Levy has written, army units were filling with unashamedly racist conscripts, and proud Jewish supremacists were ascending to become battalion leaders and commanders. “We’d barely started to experiment with this project of letting the ground troops have some kind of individual control of the intelligence gathering when we began hearing of soldiers… independently looking for things that were interesting to them and making arrests,” one reservist who developed machine learning capabilities for surveilling the West Bank in the early 2020s told me.

For example, ground troops could make a query for “M16,” which would pull up “people who maybe have an M16, maybe they are just talking about an M16—but regardless they’re going to look for them and arrest them.” What’s more, the source said, “when soldiers would go arrest someone, they could use tablets to document the people throwing stones at their vehicles. They would go back the next day and arrest those people, so it would be this endless loop.” Another 8200 veteran who surveilled the West Bank in the early 2020s echoed these concerns: “They said these systems stopped terrorism, but really it just gave troops an excuse to go out and arrest more people for something stupid.” 

It was the Hamas attacks of October 7, rather than the mounting evidence of abuses enabled by his innovations, that ended Sariel’s career. In the months after militants breached the fence between Israel and Gaza, serious lapses in intelligence protocols came to light. Stories surfaced in Israeli news media that Unit 8200 had stopped monitoring Hamas’s handheld radios and that senior commanders had classified intelligence pointing toward a possible invasion as “weak signals.”  

To some in the security establishment, Sariel’s technological experiments—which Aviv Kochavi claimed had given the IDF capabilities “akin to the movie The Matrix”—were discredited by their failure to prevent an attack of such magnitude. Retired heads of Israel’s security agencies publicly skewered the military’s obsession with emerging technologies, particularly automated translation and transcription tools, arguing that overreliance on AI had lulled the intelligence community into complacency, leading them to undervalue the human factor that makes intelligence work most effective. Such technologies were so “addictive,” the journalist Ronen Bergman wrote in September 2024, citing a former Shin Bet official, that “when it came to Gaza, the intelligence community neglected, to a certain extent, all other sources.” Other critics specifically pointed the finger at Sariel’s brash leadership. “He wanted to be a revolutionary,” the ex-security official who oversaw Sariel earlier on in his career observed, “but he wasn’t stable enough or cautious enough.”

Sariel resigned in disgrace in September 2024, claiming responsibility for the institutional failures that led to the October 7 attack. But the wave of internal criticism was not strong enough to wash away practices that had already become firmly entrenched. “The simplest thing to do for them is just to say, okay, we let the machines do it,” one reservist who served for the first few months of the Gaza war said. “It’s the easiest way for them to act, because they don’t have the capacity to do high-quality work anymore.” The AI applications Sariel had helped develop—Lavender, the Gospel, and others with equally cinematic names, like Depth of Wisdom, Alchemist, and Hunter—remained in use, leading Israeli soldiers to targets on the ground and guiding the bombs they dropped from the skies at unprecedented rates.

Today the military is rolling out ever-more-powerful software. Reservists who returned to 8200 during the war on Gaza from jobs at leading technology firms—including Alphabet, Microsoft, and Meta—began exploring how to consolidate all the information at the military’s disposal into a chatbot, which one upper-level developer described to me as a “ChatGPT” for Israeli forces. The LLM is capable of “rapidly processing large quantities of surveillance material in order to ‘answer questions’ about specific individuals,” as +972reported in partnership with Local Call and The Guardian. For example, a soldier who wanted to figure out how two individuals knew each other could query the bot to summarize their entire social history. “AI amplifies power,” an intelligence officer with knowledge of the army’s experiments with machine learning told +972. “It allows operations [utilizing] the data of far more people, enabling population control.” In late December, as the most right-wing government in Israel’s history faced accusations of war crimes and genocide, the military announced the establishment of a new artificial intelligence division that will further integrate AI into every level of the army’s decision-making. 

What all this has meant in practice for Palestinians living under Israeli occupation is hard to say for certain. Because the military keeps evidence on administrative detention cases sealed, lawyers, victims, and advocates say it is impossible to determine the precise part AI has in these charges. But given the precipitous increase in detentions overall, advocates strongly suspect that new technology is being used to ramp up arrests.  

The majority of those arrested are taken from their homes in the dead of night; armed soldiers force their way into bedrooms, blindfold and handcuff their victims, and shove them into military vans. “The effect is to terrify people,” Francis, the human rights lawyer, said. Administrative detention can drag on as long as a military court sees fit; an initial six-month sentence can be renewed again and again. Detainees can be denied clean clothes, adequate food, and medical treatment, and can be subjected to abuse, sexual assault, or torture by Israeli forces; over one hundred Palestinians have died in Israeli prisons in the last two years. The ordeal of imprisonment can derail an entire life’s trajectory. In August 2024 I interviewed a twenty-one-year-old medical school student from Jerusalem who was pulled from his bed at 3:00 AM. He spent three months in administrative detention and another five under house arrest, which forced him to drop out of graduate school and abandon his education. He was never charged with a crime.  

Meanwhile Israel and its allies are racing to deploy a new cluster of AI systems in war, to particularly devastating effect for civilians in the Middle East. The US–Israeli aerial assault on Iran rivaled the one that Israeli forces unleashed in Gaza and across Lebanon after October 7: at the height of the war the US and Israel were striking Iran roughly once every minute and a half. Both militaries advertised generative AI’s centrality to the war, broadcasting the fact that LLMs now process and summarize surveillance data in conflict zones much the same way ChatGPT condenses search engine results about, say, neighborhood bakeries. Rotem Bashi, the head of Israel’s Matzpen unit, the largest outfit devoted solely to internal IDF software development, told the Israeli website Ynet that LLMs had not only expanded Israel’s targeting capabilities but also automatically translated classified Israeli intelligence for American forces, allowing the two armies to operate in lockstep.

In mid-March the Pentagon’s chief digital and artificial intelligence officer, Cameron Stanley, stood in front of an audience at Palantir’s AIPCon and demonstrated how Maven Smart Systems, a command-and-control platform the company has created, collapsed the time it took US operatives to decide where and what to strike. Stanley navigated a computer rendering of an aerial shot of an asphalt parking lot and opened a glossy drop-down menu; the feed zeroed in on a single car in a long line of vehicles, and a text box suggested taking it out with a M2.50 caliber munition from a Browning machine gun mounted on a STRYKER armored combat vehicle. With this software, Stanley said, putting targets into a “work flow” that “closes the kill chain” was easier than ever: “left click, right click, left click.”