The Architecture of Machine-Assisted Killing
When Israeli forces began their military campaign in Gaza following the October 7, 2023 Hamas attack, they did so with a targeting apparatus unlike anything previously disclosed in the history of modern warfare. For the first time, a military had publicly — through both leaks and partial official acknowledgment — deployed an interconnected network of AI systems designed to identify, prioritize, locate, and schedule the killing of human targets at industrial scale.
The system was not a single program but an ecosystem: four distinct AI tools, each handling a different stage of the kill chain, collectively enabling a tempo of strikes that would have been impossible using traditional intelligence analysis. This report maps that ecosystem, drawing on the landmark +972 Magazine and Local Call investigation published in April 2024, subsequent reporting, and official IDF responses.
The implications extend far beyond Gaza. What Israel demonstrated — intentionally or not — is a proof of concept for machine-assisted war that every major military on Earth is now studying closely.
System 1: Lavender — The Target Generator
Lavender is an AI system trained on signals intelligence, communications data, social network analysis, and behavioral metadata to identify individuals it classifies as members of Hamas or Palestinian Islamic Jihad (PIJ). According to the +972 Magazine investigation — based on testimony from six Israeli intelligence officers who used the system — Lavender had flagged approximately 37,000 Palestinians as suspected militants by the early weeks of the campaign.
The system assigns each individual a score from 1 to 100, reflecting the AI's confidence in its classification. A score of 90 or above was treated, according to sources, as equivalent to a confirmed identification. Crucially, intelligence officers interviewed by +972 described reviewing Lavender's outputs with only cursory scrutiny — spending as little as 20 seconds per target to confirm the AI's recommendation before the target was added to a strike list.
One officer told +972: "We were not looking for errors. The machine gave the target. We just needed to make sure it wasn't a woman. That was the main check." This account, which the IDF disputed in its official response, suggests that human review had been reduced to a rubber-stamping function — legally present but operationally hollow.
"The human role was primarily to confirm that the target was male. Everything else — the identity, the affiliation, the threat level — came from the machine."
— Intelligence officer quoted in +972 Magazine investigation, April 2024Lavender's training data and algorithmic logic remain classified. The +972 report does not specify which data inputs were weighted most heavily, though officers described the system drawing on phone data, WhatsApp group membership, proximity to known militants, and other behavioral signals. The system was reportedly developed by Unit 8200, the IDF's signals intelligence and technology unit, and had been in development for several years prior to its deployment at scale in Gaza.
The error rate acknowledged within the IDF's own analysis was reportedly around 10% — meaning roughly 3,700 of those flagged may have been wrongly identified. At the casualty ratios accepted for low-ranking targets (discussed below), this margin translates to thousands of civilian deaths from targeting errors alone, before accounting for collateral damage to family members and neighbors.
System 2: Gospel — The Building Target Engine
While Lavender identifies human targets, Gospel operates one level up the abstraction hierarchy, identifying infrastructure and building targets. Gospel uses aerial surveillance, satellite imagery, and signals data to flag structures associated with Hamas military activity — command nodes, weapons storage, tunnel infrastructure, and operational meeting sites.
Officers described Gospel as having dramatically accelerated the pace at which the IDF could generate fresh target sets. Before AI-assisted targeting, the process of identifying and verifying a significant building target might take days or weeks, requiring analysts to correlate multiple intelligence streams manually. Gospel compressed this to hours or less, creating what sources called a "target bank" that could be refreshed continuously.
The scale of building destruction in Gaza — more than 60% of structures in Gaza City damaged or destroyed by February 2024 according to satellite analysis by UNOSAT — reflects in part the throughput Gospel enabled. Critics have argued that the system's classification criteria are opaque and its error tolerance, when applied to densely populated urban environments, inherently produces mass civilian casualties.
The Gospel-Lavender Interaction
Gospel and Lavender interact in a critical way: once Lavender identifies a human target, Gospel (or related systems) is used to identify that target's likely location — their home, known gathering points, or associated facilities. The strike decision then combines human identity (Lavender) with precise location (Gospel/related imagery analysis) and routes the package to Fire Factory for munitions assignment.
System 3: Fire Factory — Automated Munitions Matching
Fire Factory handles logistics and munitions selection. Given a target package — a location, a target type, a desired effect — Fire Factory cross-references available ordnance, flight windows, and aircraft availability to propose a complete strike plan. Sources describe it recommending specific bomb types and yields calibrated to the target.
Officers with direct knowledge of Fire Factory's operation described it as generating attack schedules: a queue of targets, each with recommended munitions, timing windows, and aircraft assignments. Human commanders reviewed and approved the schedule, but the computational work of matching bombs to buildings — and managing the deconfliction of simultaneous strikes — was handled by the machine.
This automation of logistics created a throughput that exceeded what Israeli air assets could historically process. One indirect indicator: in the first weeks of the conflict, the IDF conducted upwards of 1,000 air strikes in under a week — a pace requiring coordination that manual processes could not have sustained.
System 4: Where's Daddy? — Home Strike Targeting
The most disturbing system in the ecosystem, from an IHL perspective, is Where's Daddy? — a tracking tool designed to alert operators when a targeted individual has returned to their family home. The name reflects its operational logic: wait for the militant to visit his family, then strike.
Under IHL, homes are protected civilian structures unless they have been appropriated for military use. Targeting an individual in their family home — especially with a bomb large enough to destroy the structure — means knowingly killing or risking the lives of family members, including children. The +972 investigation describes explicit IDF policy for lower-ranking targets permitting the deaths of up to 100 civilians as acceptable collateral damage to eliminate a single low-ranking Hamas operative.
Where's Daddy? operationalized a doctrine of home strikes. When a target's phone signal indicated they had returned home, the system generated an alert and the target was slotted for a nighttime strike — chosen because targets were most likely to be home asleep, maximizing the probability of killing the individual. The timing, however, also maximized the probability that family members would be present.
"The 'Where's Daddy?' system was used to bomb the family home. We were told this was a legitimate way to hit Hamas members. It was accepted practice."
— Intelligence officer quoted in +972 Magazine investigation, April 2024The Four-System Kill Chain: End-to-End
Scores 2.3M Gazans, flags 37,000+ as suspected militants. Assigns probability scores 1-100. Officers confirm/reject in ~20 seconds.
Identifies buildings, command nodes, tunnel infrastructure. Generates a continuously refreshed "target bank" at machine speed.
Matches targets to available ordnance, recommends bomb types/yields, schedules strikes, deconflicts air assets automatically.
Monitors target's phone signal. Alerts operators when they return home. Enables nighttime strikes on family residences.
Pre-AI vs Post-AI Targeting: A Structural Comparison
To understand what changed, it is necessary to compare targeting practices before and after AI integration. Traditional targeting in a conflict like Gaza — as practiced in earlier Israeli operations such as Operation Cast Lead (2008) or Operation Protective Edge (2014) — was limited by human intelligence analyst capacity.
| Dimension | Pre-AI Targeting | AI-Assisted (Gaza 2023+) |
|---|---|---|
| Target identification speed | Days to weeks per significant target | Hours to minutes (Lavender) |
| Target volume | Hundreds of targets per campaign | 37,000+ human targets flagged |
| Human review time | Hours per target, multi-analyst review | ~20 seconds per Lavender output |
| Building target generation | Manual imagery analysis, days per target | Automated, continuous refresh (Gospel) |
| Munitions assignment | Manual targeting officer review | Automated matching (Fire Factory) |
| Strike tempo | Hundreds per campaign | 1,000+ per week at peak |
| Accountability | Named analysts per target decision | Diffuse — AI outputs, human "confirms" |
Collateral Damage Ratios and the Tiered Target System
Perhaps the most ethically significant disclosure in the +972 investigation concerns the explicit casualty ratios the IDF authorized for different categories of target. According to sources, the IDF operated a tiered system:
Senior Hamas commanders: Up to 100 or more civilian deaths were authorized as acceptable collateral damage. These strikes could involve large munitions dropped on populated multi-story buildings where the target was known to be present.
Mid-ranking operatives: The acceptable civilian death threshold was lower but still substantial — sources described figures in the range of 15-20 civilians per mid-level target.
Low-ranking operatives (Lavender-generated targets): Sources described a threshold of up to 15-20 civilians, with some accounts suggesting thresholds as high as 100 for certain target types. Critically, low-ranking targets were often struck with "dumb bombs" — unguided or semi-guided munitions — rather than precision weapons, increasing collateral damage.
The scale arithmetic is stark. If Lavender flagged 37,000 targets, and if even a fraction were struck under these collateral ratios, the resulting civilian death toll would run into the tens of thousands — consistent with the casualty data reported by Palestinian health authorities and cross-referenced by international monitoring organizations.
Casualty Data
By December 2024, the Gaza Health Ministry reported more than 44,000 confirmed deaths, with the UN estimating the true toll could be significantly higher once those buried in rubble are accounted for. Lancet researchers published analysis in July 2024 suggesting that indirect deaths from disease, malnutrition, and infrastructure destruction could bring the total toll to over 186,000 by August 2024 if conflict continued at the prevailing rate.
Israel disputes these figures and the methodology behind them. The IDF's official position holds that Hamas embeds in civilian infrastructure, that precautions are taken to minimize civilian harm, and that responsibility for civilian deaths rests with Hamas for its use of human shields.
International Humanitarian Law Analysis
The use of AI systems in targeting raises acute questions under IHL, specifically under the principles of distinction, proportionality, and precaution.
Distinction
IHL requires that parties to a conflict distinguish at all times between combatants and civilians. A system like Lavender — which generates probabilistic scores rather than confirmed identifications, has a self-reported ~10% error rate, and receives only 20 seconds of human review — raises serious questions about whether the legal standard of distinction is being met. If an algorithm classifies an individual as a combatant and a human rubber-stamps that classification in 20 seconds without independent verification, the legal accountability chain has been functionally severed.
Proportionality
IHL prohibits attacks expected to cause civilian harm that is excessive relative to the anticipated military advantage. The disclosed collateral damage thresholds — up to 100 civilians per low-ranking operative — represent explicit calculations that many international law scholars argue exceed what proportionality permits. Several United Nations Special Rapporteurs have characterized aspects of the campaign as potentially constituting war crimes, independent of the AI question.
Precaution
IHL requires parties to take all feasible precautions to avoid or minimize civilian harm. Critics argue that the 20-second review window and the "dumb bomb" policy for Lavender-generated targets represent the active abandonment of precaution. If a higher standard of review were applied, the strike tempo would necessarily decrease — which suggests the 20-second window was a deliberate operational choice to maintain throughput, not a minimum imposed by circumstance.
The IDF's Official Position
The IDF disputed several key claims in the +972 report. In a statement, the IDF said:
"The IDF does not use an artificial intelligence system that identifies terrorist operatives or that recommends targets. The IDF uses various tools to cross-reference and identify terrorist operatives... Hamas terrorists bear responsibility for the casualties in the Gaza Strip."
— IDF spokesperson statement in response to +972 Magazine investigation, April 2024The IDF's denial is notably specific in scope — denying that AI "recommends targets" — while not addressing whether AI systems like those described are used for other functions in the targeting pipeline. The framing that AI assists with "cross-referencing" rather than "recommending" is a definitional distinction that critics argue is immaterial to the legal and ethical analysis.
Subsequent reporting by The Guardian and other outlets citing additional sources broadly corroborated the +972 findings. The New York Times published its own investigation reaching similar conclusions about the scale and nature of AI-assisted targeting in the conflict.
Precedent-Setting Implications
The Gaza conflict has established several precedents — whether or not they were intended as such — that will shape the future of autonomous and AI-assisted warfare globally.
1. Scale of AI-generated targeting: No prior conflict had seen AI systems flag tens of thousands of humans for potential lethal action. The normalization of this scale, absent international legal pushback, sets a floor for future conflicts.
2. The rubber-stamp review problem: The 20-second review window demonstrates that "meaningful human control" — the standard proposed by advocates for responsible autonomous weapons — can be present in form while being absent in function. Policymakers working on autonomous weapons treaties must grapple with this gap.
3. Explicit collateral ratios: The disclosed practice of pre-authorizing specific civilian death numbers per target class represents a level of systematization of civilian harm that, if replicated, would fundamentally transform the calculus of urban warfare.
4. Home strike doctrine: The Where's Daddy? system operationalizes a doctrine of targeting individuals in civilian residences. If accepted as precedent, this doctrine would eliminate the protection historically afforded to homes under IHL.
5. AI accountability gap: When targeting decisions flow from AI outputs confirmed by humans in 20-second windows, the traditional accountability structures of military command — where a named officer takes responsibility for a targeting decision — become functionally inapplicable. This accountability gap has no established remedy in international law.
Global Military Response
Defense establishments in the United States, United Kingdom, China, France, and Australia have all indicated interest in the capabilities demonstrated in Gaza. The US Department of Defense's Project Maven — which uses AI to analyze drone footage — represents a parallel development track. DARPA's autonomous targeting research programs have accelerated timelines following the conflict.
The key lesson militaries are drawing is not necessarily that the specific systems should be replicated, but that the integration model — AI flagging, human rubber-stamping, automated logistics — represents a viable operational architecture. The controversy around the collateral ratios and the +972 disclosures is treated by most defense establishments as a disclosure problem rather than a capability problem.
China's PLA has been particularly attentive. Its published military AI doctrine emphasizes "intelligentized warfare" (智能化战争) and calls for AI integration across all phases of the kill chain. The Gaza model, flawed as it is legally, provides the first real-world validation that such integration is operationally feasible at scale.
What International Law Must Confront
The Gaza AI targeting case arrives at a moment when international negotiations on autonomous weapons systems have stalled. The Convention on Certain Conventional Weapons (CCW) process at the United Nations has been discussing lethal autonomous weapons systems (LAWS) since 2014 without producing a binding instrument. The Gaza disclosures inject new urgency — and new complexity — into that debate.
The core challenge is that the systems described are not "fully autonomous" in the science-fiction sense of a robot deciding independently to kill. They are human-machine hybrids in which human oversight has been compressed to near-zero without being formally removed. Existing proposals for LAWS governance struggle to address this architecture, because it technically preserves human decision-making while gutting its substance.
Scholars of international humanitarian law have proposed several responses: minimum review time standards, mandatory transparency reporting on AI use in targeting, prohibition on AI systems trained on datasets without independent verification, and criminal accountability standards that address rubber-stamp approval chains. None of these proposals currently has sufficient state support to advance.
Key Takeaways
- Lavender flagged 37,000+ targets using behavioral data and social network analysis, with officers reviewing outputs in approximately 20 seconds.
- Gospel automated building target generation, creating a continuously refreshed "target bank" that enabled unprecedented strike volume.
- Fire Factory handled munitions matching and scheduling, automating the logistics of which bomb hits which target at what time.
- Where's Daddy? tracked targets to their family homes and triggered nighttime strikes when they returned — a doctrine that knowingly put civilians at risk.
- Collateral damage thresholds were pre-authorized at up to 100 civilian deaths per low-ranking militant, according to +972 investigation sources.
- The IDF disputed the characterization of these systems as target-recommenders while not fully denying their existence or use.
- No international legal framework currently governs AI-assisted targeting at this scale or with this architecture.
- Every major military is studying this model, making the legal and ethical precedents set here likely to propagate globally.
Sources and Further Reading
+972 Magazine / Local Call: "Lavender: The AI machine directing Israel's bombing spree in Gaza" — Yuval Abraham, April 3, 2024
+972 Magazine: "A mass assassination factory: Inside Israel's calculated bombing of Gaza" — Yuval Abraham, Meron Rapoport, November 30, 2023
The Guardian: "Israel used AI to identify targets in Gaza, intelligence sources say" — November 2023
UN Office for the Coordination of Humanitarian Affairs: Gaza Situation Reports, 2023-2024
UNOSAT: Building Damage Assessment, Gaza Strip, 2024
The Lancet: "Counting the dead in Gaza: difficult but essential" — July 2024
Human Rights Watch: "Gaza: Israeli Strikes Killed Families" — 2024
IDF Spokesperson: Official statement in response to +972 Magazine investigation, April 2024