The approve button¶

The makers and users of the targeting systems on which the United States, Israel and NATO rely all say a person stays in the decision. What varies is how much has already been decided by the time that person is asked.

A standing programme¶

On 9 March 2026 the US deputy secretary of defense, Steve Feinberg, wrote to senior Pentagon leaders and commanders that Palantir’s Maven Smart System “will become an official program of record”, with oversight moving to the Pentagon’s Chief Digital and AI Office. Reuters reported the memo on 20 March, describing Maven as “a command-and-control software platform that analyzes battlefield data and identifies targets”, already the primary AI operating system of a military that had carried out thousands of targeted strikes against Iran in the previous three weeks. Maven reads data from satellites, drones, radars and intelligence reports; the problem it answers, CSIS notes, goes back to wars in which there were more cameras recording video than trained analysts to watch it. Palantir says its software does not make lethal decisions, and that humans remain responsible for selecting and approving targets.

The decision to make it a standing programme came while the military it serves was busy. By September a Pentagon official was counting “13,000 targets in 38 days”.

Twenty seconds¶

On 3 April 2024 Yuval Abraham, writing for +972 Magazine and Local Call, published the accounts of six Israeli intelligence officers about an Israeli army system called Lavender, which, they said, “clocked as many as 37,000 Palestinians as suspected militants”. One officer described the human check: “I would invest 20 seconds for each target at this stage.” The army’s spokesperson denied the existence of a kill list and called its AI tools auxiliary, with independent review by analysts required. Lavender is the army’s own system, not Palantir’s.

A person approved each target. The sources put the human check at about twenty seconds a target.

The last confirmation¶

In May 2026 Palantir representatives told Janes that the platform under Maven can be set up so that an operator approves a target and “the system connects the firing circuit from sensor to effector”. In their example, a message is sent to an aircraft’s fire control computer, and “a pilot would only need to confirm the engagement by releasing the missile”. Palantir’s UK defence lead, Polly Scully, said a human in the loop remains, and that the system is “adaptable to the requirement of the military, to the extent that it [the military] wants to automate targeting”, within legal bounds.

In the example two people act, one approving the target and one releasing the weapon, and the system does the joining between them.

The alliance’s copy¶

NATO’s communications agency finalised its purchase of a NATO version of Maven on 25 March 2025, for “intelligence fusion and targeting, battlespace awareness and planning, and accelerated decision-making”. On 22 June 2026 it reached what NATO calls full technical operational capability, Janes reported, accredited for classified networks, with “NATO data on NATO servers in NATO data centres”.

In the Netherlands the government told parliament in May 2026 that several operational defence units use Palantir software, “onder strikte maatregelen en in volledige afgesloten omgevingen”, as a trial. A week later Defence said it wants to replace it with a European alternative within two years; the state secretary, Derk Boswijk, would not advise cutting the ties at once. Written answers to parliament named no end date, only alternatives to explore and dependencies to reduce.

The servers are the alliance’s own; the way of deciding came with the software.

Coming home¶

In September 2026 VPRO Tegenlicht took up the idea, from the journalist Rania Khalek, of an imperial boomerang: “geweld, technologie en controle”, violence, technology and control, tried out far away and brought back as border control and predictive policing. A short from the same programme was titled Big Tech test dodelijke AI op burgers, Big Tech tests lethal AI on civilians. Nothing on file shows these targeting systems turned inward; the boomerang is Tegenlicht’s argument about where such tools have gone before.

Each of these systems still asks a person to approve. Only the people doing the approving have said how long it took, twenty seconds, and the makers say a person approves without saying what that person sees, or what is left to decide by the time the button is offered.