AI in War: What Is Documented, Claimed and Disputed

Software now helps decide who gets bombed. That sentence would have sounded like science fiction a decade ago; in 2026 it is the documented core of two active wars and a stalled diplomatic fight in Geneva. The hard part is separating what is confirmed from what is claimed by anonymous sources and what militaries actively dispute. This piece tries to hold that line.

From a handful of targets a year to hundreds a day

Israel's military has publicly acknowledged an AI system called "the Gospel" (Habsora) that helps generate strike targets. Reporting by NPR and others describes it fusing intercepted signals, imagery and open-source data to recommend infrastructure targets. Israeli officials said it helped produce targets far faster than human analysts alone — figures around 100 a day have circulated, against roughly 50 a year in Gaza previously. That the Gospel exists is confirmed; the exact throughput and accuracy are claimed by the military and not independently verified.

Far more contested is "Lavender." In April 2024, +972 Magazine and Local Call reported, citing six anonymous Israeli intelligence officers, a system that scored Gaza residents on the likelihood of being militants and flagged up to 37,000 people. The same sources described permissive collateral ratios — reportedly 15–20 civilians tolerated per junior operative. The Israeli military denies keeping any "kill list," calling such tools "auxiliary" aids that still require an analyst's independent check. So: the reporting is real and detailed; the specific numbers and the thinness of human review are disputed, and rest largely on anonymous testimony.

Where the human sits in the loop

The debate turns on one distinction. In a human-in-the-loop system, a person approves each engagement. On-the-loop means a human supervises and can abort but does not authorise every shot. Out-of-the-loop — often called a lethal autonomous weapon — means the machine selects and strikes with no one intervening.

Most systems described in Gaza and Ukraine keep a human somewhere in that chain. The worry experts raise is not killer robots roaming free but "automation bias": when a screen outputs a name and a confidence score under time pressure, meaningful human review can erode into a rubber stamp. The label "human oversight" can mask how little judgement is actually exercised.

Cheap drones and the drift toward autonomy

Ukraine has become the world's live laboratory for AI-assisted strike drones. Analysts at CSIS and defence reporters describe first-person-view drones using off-the-shelf machine vision — software that locks onto a target in the final seconds, helping the drone hit even when the radio link is jammed. This is terminal guidance, not free-roaming autonomy: a human still picks and launches. Russia has fielded its own loitering munitions with reported autonomous features. Both sides are pushing toward less human input, driven by electronic jamming that severs the operator's link.

The clearest precedent for fully autonomous engagement remains disputed. A 2021 UN Panel of Experts report on Libya described a Turkish-made STM Kargu-2 loitering munition that may have attacked fighters in 2020 with no operator connection — "fire, forget and find." It was widely called the first autonomous strike on humans. The manufacturer disputes this, insisting its drones operate under a man-in-the-loop principle, and the report never confirmed casualties. It is a warning sign, not a settled fact. Our look at AI versus humans across five real tasks shows why machine perception still fails in exactly the edge cases that matter most on a battlefield.

The regulation fight is real but stuck

Governments have discussed lethal autonomous weapons at the UN's Convention on Certain Conventional Weapons since 2014 — and, as of 2025, were still holding consultations rather than treaty negotiations. Momentum is building elsewhere. In November 2025, the UN General Assembly's First Committee adopted a resolution on autonomous weapons by 156 votes to 5, with 8 abstentions. Secretary-General António Guterres called such systems "politically unacceptable, morally repugnant" and, with the ICRC, urged a legally binding instrument by 2026.

The ICRC's position is specific: prohibit unpredictable autonomous weapons and those designed to target people, and strictly restrict the rest. A ban is not imminent — major military powers resist binding limits — but the diplomatic direction is clear.

What the experts actually fear

Three concerns recur. Accountability: if a flawed algorithm contributes to an unlawful strike, who answers — the coder, the commander, the vendor? Existing law was written for human decisions. Error and bias: a model trained mostly on confirmed targets has little sense of what a non-target looks like, and speed multiplies any mistake. Escalation: machines reacting to machines at superhuman speed could tip a crisis into war before humans can intervene.

None of this requires a robot uprising to be dangerous. The risk is quieter: accountability draining out of lethal decisions one confidence score at a time. This is regulation catching up to a technology already deployed — read alongside the EU AI Act's approach to high-risk systems, which pointedly exempts military use, leaving the hardest cases to a treaty that does not yet exist.

The takeaway: AI is already in the kill chain, the humans are often still there but doing less, and the rules meant to govern it are years behind the hardware.