In the weeks after Hamas launched its October 7, 2023 attack, the Israeli military deployed a constellation of AI-assisted targeting tools unlike anything previously confirmed in combat. One system — an internal machine learning tool called Lavender — would generate a kill list of over 37,000 people before the ground offensive had fully begun.
The Investigation That Broke the Story
The existence of Lavender and three related systems was first revealed in April 2024 through a joint investigation by +972 Magazine and Local Call, two Israeli news outlets known for their access to IDF sources. The reporting drew on interviews with seven current and former Israeli intelligence officers who had direct knowledge of or involvement with these systems during the early months of the Gaza campaign. The investigation represented one of the most detailed confirmations of AI use in a live military targeting chain ever made public.
The IDF neither confirmed nor denied the specific mechanics described. In a statement, the military said it "does not use an artificial intelligence system that identifies terrorist operatives or tries to predict whether a person is a terrorist" and that "all targets are thoroughly verified by qualified intelligence analysts." The officers interviewed by +972, however, described a starkly different operational reality.
How Lavender Worked
Lavender was a machine learning model trained on data profiles of known Hamas and Palestinian Islamic Jihad (PIJ) members. According to the intelligence officers interviewed, the system analyzed signals across a range of data sources: phone metadata, messaging application usage, social network associations, family ties to known operatives, movement patterns, and association with other flagged individuals. It assigned each male in Gaza above a certain age a score from 1 to 100 representing its assessed probability of that individual being a militant.
At the beginning of the post-October 7 campaign, the system had been trained on a relatively limited dataset and its outputs carried significant uncertainty. One officer described the system as having an error rate of roughly 10 percent — meaning that for every ten people Lavender flagged, one was likely not a militant at all. In a kill list of 37,000 people, that translates to approximately 3,700 individuals flagged as legitimate targets who were, by the system's own internal accuracy estimate, innocent civilians.
"The machine did 80 percent of the work. We basically approved everything. There was no way to do 37,000 human investigations in that timeframe."
— IDF intelligence officer, quoted by +972 Magazine, April 2024The Lavender list was not the final word. It fed into broader operational planning that involved additional automated systems and — in theory — human review. But the nature of that review changed dramatically under wartime conditions.
The Gospel and Fire Factory
Lavender operated alongside at least two other AI-assisted targeting systems. The Gospel, which the IDF had publicly discussed in limited terms before October 7, was designed to generate targets related to infrastructure: military compounds, tunnel shafts, rocket production facilities, and the buildings associated with Hamas command functions. The system automated what had previously taken weeks of intelligence work, reportedly capable of producing dozens of new infrastructure targets per day at the campaign's peak.
Fire Factory was the logistics layer. Once Lavender or The Gospel identified a target, Fire Factory would calculate the recommended munitions load, match available ordnance to the target's assessed hardness, schedule strike windows, and produce preliminary estimates of expected collateral damage. Officers described the system as handling what would previously have been intensive work by munitions and targeting specialists, compressing that process to near-instantaneous outputs.
Together, the three systems created an automated pipeline from suspect identification to strike scheduling that could process thousands of potential targets faster than any pre-AI Israeli targeting operation. The speed of the system was, by design, one of its core military capabilities. But it also fundamentally altered the nature of human involvement in lethal decisions.
Where's Daddy: The Home Strike Doctrine
A fourth system, referred to in the +972 reporting by the informal name "Where's Daddy?", tracked when a flagged individual returned to their private residence. The tactical logic was deliberate: bombing a target at home rather than in a tunnel or military facility reduced the risk of casualties to Israeli soldiers. Strikes on residential buildings were, in this framework, not collateral to the military operation but central to it.
This approach had a direct and documented effect on civilian casualty rates. When a flagged individual of lower rank — someone assessed by Lavender as likely a low-level Hamas operative rather than a senior commander — returned home, strikes were authorized that could be expected to kill family members as a byproduct. Intelligence officers described explicit "collateral damage ratios" that defined how many civilian deaths were acceptable per target based on the target's assessed military value.
For a senior military commander, the accepted collateral damage figure was reportedly up to 100 civilian deaths. For a low-ranking militant — the majority of Lavender's 37,000 flagged individuals — the accepted figure was between 15 and 20 civilian deaths. Officers described family homes as deliberately targeted at night, when subjects were likely to be home, maximizing strike success against the individual while simultaneously ensuring civilian family members would also be present.
The 20-Second Approval Window
The compressed human oversight timescale reported by +972 is perhaps the most operationally significant detail in the investigation. During the early phase of the campaign, intelligence officers described spending approximately 20 seconds reviewing a Lavender-generated profile before approving it for a strike authorization. This was not described as an unofficial shortcut but as the practical consequence of scale: with thousands of targets moving through the system, extended review of each was operationally impossible.
The 20-second window typically involved an officer verifying that the Lavender-flagged individual was male and confirming the associated residential address. The underlying AI determination — that the individual was likely a militant — was treated as a given, not an item for independent scrutiny. Officers described the review as a "rubber stamp" in practice, even if it formally preserved the fiction of human decision authority in the targeting chain.
This pattern is exactly what AI ethics researchers and international humanitarian law scholars have described as the risk of "automation bias" — the tendency of human operators to over-trust machine outputs, particularly under conditions of time pressure and high operational tempo. The Lavender disclosures provided, for the first time, a documented real-world instance of that phenomenon operating at scale in active combat targeting.
Pre-AI Targeting: A Comparison
Before AI-assisted targeting tools were deployed, Israeli intelligence targeting operations in Gaza followed a more labor-intensive process. A senior Hamas commander might require weeks of collection, analysis, and multi-source corroboration before being placed on an approved target list. A low-level operative might not be targeted individually at all — the intelligence investment required to confirm and locate such individuals made them an inefficient use of targeting resources.
AI targeting systems eliminated that constraint. By automating the analysis phase and dramatically lowering the cost of target generation, Lavender and its companion systems enabled the IDF to target individuals at a tier of the Hamas organizational structure that had previously been practically unreachable in anything approaching real time. The same logic that made the system militarily efficient — its ability to process thousands of targets with minimal human analyst time — also meant that the error rate of the AI, applied across a list of 37,000 people, translated to errors measured in the thousands.
International Humanitarian Law Implications
The Lavender disclosures raised immediate and pointed questions under international humanitarian law (IHL), specifically the principles of distinction, proportionality, and precaution as codified in the 1977 Additional Protocols to the Geneva Conventions.
The principle of distinction requires parties to a conflict to differentiate between combatants and civilians and to direct attacks only against military objectives. A targeting system with a self-assessed 10 percent false positive rate — operating at scale and with minimal human review — creates a structural mechanism for violating this principle, regardless of operator intent.
The proportionality rule prohibits attacks expected to cause civilian casualties that are excessive in relation to the anticipated military advantage. The explicit "collateral damage ratios" described in the +972 investigation — including the reported acceptance of 15 to 20 civilian deaths for a low-ranking operative — constitute precisely the kind of proportionality calculus IHL requires, but that calculus, as described, produced outcomes that many legal scholars consider incompatible with the law's requirements in the context of targeting junior operatives in civilian residential structures.
Precaution requirements demand that parties take all feasible steps to verify targets and minimize civilian harm. A 20-second human review window, against a machine-generated list that the operators themselves acknowledged was not individually scrutinized, raises serious questions about whether the required "feasible precautions" were taken in each strike authorization.
Human Rights Watch, Amnesty International, and numerous academic international law experts issued statements in 2024 calling for independent investigations into the AI targeting disclosures. The UN Special Rapporteur on extrajudicial killings specifically cited the +972 reporting in a report to the Human Rights Council, calling the described practices "deeply troubling" and potentially unlawful.
IDF Responses and Justifications
The IDF's official response to the +972 investigation was consistent denial of the specific operational details described. Military spokespersons maintained that all strikes were conducted in accordance with IHL, that human officers retained full decision authority in all targeting actions, and that the military does not use systems that automatically designate individuals as targets without human oversight.
Unofficially, IDF-adjacent commentary and some military analysts offered a different framing: that the scale and speed of the threat posed by Hamas after October 7 required a corresponding operational tempo, that AI-assisted targeting reduced the total risk to Israeli personnel, and that the alternative to rapid AI-assisted targeting was either accepting greater Israeli casualties or foregoing time-sensitive strikes against an adversary actively embedded in the civilian population.
This framing represents a consequentialist argument for AI targeting efficiency that international law scholars generally reject as a basis for overriding IHL protections. The laws of armed conflict do not contain a military necessity exception to the distinction and proportionality principles — an argument the Israeli military itself has historically used against adversaries in other contexts.
Casualty Statistics and the Weight of Scale
By early 2025, the Gaza Ministry of Health — whose figures are generally accepted by the United Nations as the most reliable available — reported more than 46,000 Palestinian deaths. Independent analyses by organizations including Airwaves, the Humanitarian Data Exchange, and various academic research teams found that a disproportionate share of casualties were women, children, and elderly — demographics inconsistent with what would be expected if strikes were predominantly targeting Hamas combatants.
The Lavender disclosures provide a structural explanation for patterns in the casualty data that are otherwise difficult to account for: a targeting system that flagged 37,000 individuals, applied collateral damage ratios that explicitly accepted civilian deaths per strike, and prioritized home strikes at night — when families are present — would predictably produce the kind of civilian casualty distribution documented in the conflict data.
No independent investigation has yet been completed that directly links specific strikes to Lavender-generated targeting decisions. Israel has not permitted independent international investigators access to the relevant military data. But the structural logic of the system, as described by its operators, maps closely onto the observable pattern of casualties on the ground.
- Lavender was a machine learning model that flagged more than 37,000 Palestinians as potential Hamas or PIJ operatives, generating what amounted to an AI-compiled kill list.
- Human review of each target was reduced to approximately 20 seconds in practice, representing near-total reliance on machine outputs rather than independent human judgment.
- The Gospel automated infrastructure targeting; Fire Factory automated munitions matching and collateral damage calculation; "Where's Daddy?" tracked targets to residential addresses for strikes.
- Explicit collateral damage ratios accepted 15 to 20 civilian deaths per low-ranking target, and up to 100 per senior commander.
- International humanitarian law scholars have identified the described practices as potentially violating the principles of distinction, proportionality, and precaution under the Geneva Conventions.
- The Gaza campaign represents the most extensively documented deployment of AI targeting systems in active combat in history, with profound implications for the laws of war and for future conflicts.
The Precedent Problem
Whatever legal conclusions are eventually drawn about the Gaza campaign, the operational precedent established by Israel's AI targeting architecture is now part of the global military record. The demonstrated capacity to generate thousands of individualized strike targets with machine speed — and to compress human decision authority to near-zero while formally preserving it — represents a qualitative shift in what is operationally possible in urban counterterrorism warfare.
Every major military that has observed the Gaza campaign closely is now analyzing what was used, how it worked, and how to build equivalent capabilities. China's People's Liberation Army, the United States military, and Russia's defense establishment all have active programs in AI-assisted targeting. The question is no longer whether autonomous and semi-autonomous targeting will spread — it is how quickly, in what form, and whether any international legal framework will prove capable of governing its use before the next major urban conflict puts another AI targeting stack into operation.
The Lavender system may be remembered less for what it did in Gaza than for what it demonstrated to every military on earth: that AI can compress a targeting cycle from weeks to seconds, that human oversight can be reduced to a procedural formality while remaining formally intact, and that the laws of armed conflict as currently written offer no clear technical threshold at which an AI targeting system becomes legally impermissible. That gap — between what is now technologically possible and what international law was designed to prevent — is the most consequential weapons policy question of the current decade.