Airwars and AI Now dissect the AI-run military kill chain

A visual investigation out this week from Airwars and the AI Now Institute walks through something most coverage of military AI never quite gets to, which is what the machines are actually doing at each step. Anatomy of an AI Kill Chain, launched on July 28, 2026, breaks the process into six stages: sensor and data gathering, surveillance, intelligence and identification, target selection, strike, and post-strike assessment. The authors, Sophia Goodfriend, Heidy Khlaaf, Namir Shabibi, Joe Dyke and Nathan Walker, use a fictitious kill chain to show how AI now mediates each of those stages, because the real proprietary systems are largely censored and secret.

The number that stuck with me is inside the report and picked up by The Register: in some operations only two of the six stages of the US military kill chain still involve humans in the loop, with a third under some human oversight and the rest fully automated. Whether or not you accept that as a settled figure, the direction is clear enough that Airwars, which describes itself as a not-for-profit transparency watchdog, felt it needed a public explainer at all.

The part of the argument the authors lean on hardest is not the tech, it is the human reviewer. Goodfriend describes "the pernicious problem of automation bias, especially under the time constraint of war," and Khlaaf is blunter, saying oversight becomes "really superficial in practice" and that human beings "just become rubber stamps at that point." The failure modes the piece points at are the familiar ones from civilian AI: translation errors, computer vision misidentification, GPS and jamming effects on neural networks, algorithmic risk scoring off social media. They just carry very different consequences at the end of a kill chain.

The honest caveat is that this is a fictional composite, not a leaked target file. The reporting does not name the specific vendors, models or contracts that sit inside real operations, and it does not tell you which militaries the "two of six" figure covers beyond the US context. What it does give you is a taxonomy sharp enough that procurement officials, auditors and journalists can start asking stage-by-stage questions instead of arguing about "AI in war" as one blob. That framework is the thing worth taking forward.