20 Seconds. That Is the Time an AI Now Gets to Decide Who Lives
In the early weeks of the Gaza war, an artificial intelligence system known as Lavender marked 37,000 Palestinians as suspected militants. The officers reviewing each name, according to six Israeli intelligence personnel who spoke to +972 Magazine and Local Call, spent about 20 seconds per target before authorising a bombing. Their review, in many cases, was just to confirm that the system had selected a male.
The system was known to be wrong, by its own internal measure, in 10 per cent of cases.
One officer described his role as a rubber stamp on the machine's decisions.
That phrase captures the moment a doctrine quietly changed. The phrase human in the loop survived. The function it described did not.
Source: +972 Magazine investigation, April 2024
What the Systems Actually Do
The +972 reporting describes three connected AI systems.
Lavender assigns each adult male in Gaza a numerical score from 1 to 100 based on cell phone records, social media activity, and location data. Higher scores indicate greater estimated likelihood of affiliation with Hamas or Palestinian Islamic Jihad. The output is a list of human targets, generated faster than any traditional intelligence process could produce them.
Where's Daddy tracks individuals on that list and sends an automatic alert when they enter their family home. The bombing is then authorised on that home, while the target is inside it, with the family. The Gospel, or Habsora, performs the same function for buildings.
According to the +972 sources, the IDF authorised the killing of up to 15 to 20 civilians as collateral damage for each low-ranking militant marked by Lavender, and over 100 civilians for each senior commander. Less precise dumb bombs were preferred for lower-ranked targets, to preserve precision munitions for senior figures.
The IDF has confirmed it uses AI as a decision support tool but disputes the +972 characterisation of human review as a rubber stamp. The technical architecture itself is not in dispute.
Source: Verfassungsblog legal analysis
The Three Problems Every Military Will Now Face
Automation bias. When a computer presents a confident answer, humans tend to accept it, even when they know the system has a measurable error rate. A 90 per cent confidence figure feels high. It also means that for every 37,000 names, roughly 3,700 are wrong. An officer with 20 seconds has no realistic way to identify which 3,700.
The responsibility gap. When an AI recommends a target and a human approves it, who is responsible if the target turns out to be a civilian? The officer who clicked? The commander who set the threshold? The engineer who trained the model? International humanitarian law was not written for this question. Every military that fields these systems is making policy by deployment, not by deliberation.
Speed versus judgement. The promise of these systems is speed. The cost is that human review must run at the same speed, or oversight becomes ceremonial. Twenty seconds is not the symptom. It is the design.
The Choice the World is Making Right Now
The Pentagon has built parallel capabilities through Project Maven and, in late 2025, signed contracts worth up to 200 million dollars each with Google, Microsoft, OpenAI, and Anthropic for frontier AI services.
In February 2026, Anthropic publicly drew a line. The company refused to allow its Claude models to be used for fully autonomous lethal targeting, prompting public pushback from US Secretary of Defense Pete Hegseth. It was the first time a major AI developer named the autonomous targeting use case and refused it.
At the United Nations, the Group of Governmental Experts on Lethal Autonomous Weapons Systems had a 2026 deadline to produce a binding treaty. That deadline has passed. No treaty exists.
Why This Matters for India
India is not yet deploying anything publicly comparable to Lavender. But India is building the foundational layer. Indian air command and control systems are integrating AI. Predictive intelligence tools are being deployed along contested borders. The sovereign AI programme, with SarvamAI and BharatGen, is being positioned to provide indigenous foundation models for defence applications.
The 20-second question will arrive on Indian operations desks within five years. The doctrine that decides the answer should be written now, before the system arrives, not afterwards when an officer is sitting at a console with a queue of 200 targets and a commander asking why approvals are running slow.
What is the minimum human review time below which a strike cannot be authorised? What is the maximum acceptable false positive rate? What is the collateral damage threshold per target category, and who signs that document? What does the Indian armed forces actually want a human in the loop to do, in plain operational language, when an AI is generating targets faster than humans can read them?
Twenty Seconds
The 20-second figure is not the scandal. It is the natural endpoint of a system designed for speed over deliberation. The scandal, if there is one, is that the doctrine that allowed it was never publicly written down. It evolved through use, until one day six officers sat down with a journalist and described what their job had become.
Twenty seconds is long enough to scratch your nose. Long enough to read a sentence and a half. Long enough to confirm that a name on a screen belongs to a man.
It is not long enough to decide whether that man, and his wife, and his three children, should die in their home tonight.
Every modern military will soon face the choice of whether to build systems that force this question, or systems that quietly answer it. India will face the same choice. The question is whether the answer arrives by doctrine, by debate, and by law, or whether it arrives the way it arrived for Lavender. By deployment.