There is a new paper of Rainer Rehak and Taylor Kate Woodcock. “Automating civilian harm: On Israel's use of the AI-enabled targeting system Lavender in Gaza and International Humanitarian Law. 2026 ACM Conference, 10.1145/3805689.3812357” that has some cruel details expanding on the current Wikipedia entry on AI-assisted targeting in the Gaza Strip.
The information depends on investigative journalism, built on anonymous IDF sources and leaked documents, but not independent forensic access. So there may be some limitations but this is all we know. The IDF’s response is consistently reported as a partial denial: the IDF stated some claims were baseless while others reflected a flawed understanding of IDF directives, describing a human-controlled process. There are three distinct systems.
- Lavender is an “AI-powered database” which lists tens of thousands of Palestinian men linked by AI to Hamas or Palestinian Islamic Jihad, used for target recommendation
- Gospel automatically reviews surveillance data looking for buildings and equipment thought to belong to the enemy and recommends bombing targets. One officer told the Guardian “I would invest 20 seconds for each target at this stage, and do dozens of them every day. I had zero added-value as a human, apart from being a stamp of approval.”
- Where’s Daddy? is designed specifically to strike people at home with family present, because it was operationally easier than catching them during military activity.
Rehak and Woodcock now write
The digital infrastructure necessary for collecting, storing, and processing the data is provided by large US companies. The cloud storage of approximately 13.6 petabytes (13,600 terabytes) is provided by Google and Amazon under the name Project Nimbus. Cisco, Dell and Red Hat/IBM provide additional IT services. Palantir is responsible for integrating data from heterogeneous sources, and the AI applications are provided by Microsoft/OpenAI and Google; however not without internal resistance. To ensure optimal collaboration between the companies and the Israeli military, there are various personnel overlaps, e.g. Microsoft employees are also part of Unit 8200 of the IDF or regularly switch roles.
Their central finding is that the systems fail all three by design, not by accident. Reported configuration details they cite: thresholds for what counts as a “target” are adjusted up or down purely to hit a daily quota; pre-set tolerances allow roughly 15 civilian deaths per “junior” target and far more for “senior” targets; cheap, imprecise munitions are deliberately used on lower-value targets; and human review of each AI-generated target reportedly takes around 20 seconds, often just confirming the target’s sex. Internal IDF checks reportedly found about a 10% misidentification rate by the military’s own loose criteria.
The authors argue this combination amounts to a “reckless disregard” for distinction (people are targeted on statistical resemblance, not verified conduct), a failure of precautions (known error-prone, brittle systems are used uncritically at speed and scale), and very likely disproportionate harm given the resulting civilian death toll. They conclude that “targeted killing” is a misleading label for what they characterize as a more indiscriminate, quota-driven process, and that - given Israel’s technical sophistication - the high civilian toll is more plausibly a deliberate policy choice than an unintended side effect.
The cynical naming convention (Lavender, Gospel, Where’s Daddy?) sits awkwardly next to the lethality the AI systems used.
As if that weren’t enough, there seems to be another system not mentioned by Rehak and Woodcock called “Server in the Sky” (SITS) on drone fleets attributed to Haaretz reporting earlier this month. It deployed across Hermes 450 and 900 drones, designed to process intelligence and detect/classify targets in real time.
According to the documents, the algorithm independently analyzes the intelligence gathered by the drones’ sensors and cameras, automatically detecting targets, classifying them and deciding whether to track them or pass them on - to the command center, air force pilots or troops on the ground.The server and the analytics it runs also allow the drone fleet to be managed autonomously, handing over tasks as the drones surveil a defined sector, shifting the burden among these unmanned aircraft to maintain continuous visibility.
So far neither ICJ (International Court of Justice) nor ICC (International Criminal Court) has so far issued a ruling that evaluates the Lavender/Gospel/Where’s-Daddy systems specifically as evidence – the charges so far center mainly on starvation as a weapon, destruction of infrastructure, and the overall casualty count.
In Terror (2015), von Schirach has a pilot decide whether to shoot down a hijacked airliner heading for a packed stadium, killing 164 passengers to save tens of thousands. Germany’s Federal Constitutional Court had already ruled this unconstitutional: such a shoot-down violates the inviolable human dignity (Art. 1 GG) and right to life (Art. 2 GG) of the innocent passengers – a person may never be reduced to a mere calculation. It’s an absolute prohibition, not a proportionality balance; you simply don’t weigh lives against lives. That’s the sharp contrast with the IDF system that does the opposite. It pre-configures civilian deaths as tolerable collateral (15 for “junior,” hundreds for “senior” targets), not even as an emergency judgment under pressure like von Schirach’s pilot, but as a routine, repeatable daily setting.
Why does the reporting on the AI systems circulates only in NGO and academic analyses but has neither reached the public nor the ICJ?