Airwars AI kill chain report maps automation in battlefield targeting

Airwars has published an AI kill chain report that maps how machine-learning systems can influence military targeting, from data collection to post-strike assessment. The nonprofit transparency watchdog says the work is meant to show how decisions about surveillance, target selection and killing can be spread across a stack of automated systems rather than a single autonomous weapon.

The investigation, titled Anatomy of an AI Kill Chain, uses a fictional targeting sequence to illustrate processes that real militaries use to find and attack targets. Airwars said the fictional framing was chosen because current military systems are often proprietary, classified or censored, making it difficult to reconstruct a live operation without overstating what is known.

What is the Airwars AI kill chain report?

A kill chain is the sequence of steps a military follows to identify, approve, attack and review a target. Airwars breaks that sequence into six stages: decision support for data gathering, surveillance, intelligence and identification, target selection, strikes, and post-strike assessment.

The report argues that AI can appear at each of those points. It cites a recent book on US military AI that said some operations now involve humans in only two of the six stages of the US military kill chain, with partial human oversight in a third and the rest fully automated.

Sophia Goodfriend, a research fellow at the University of Cambridge’s Pembroke College and a co-author of the report, told The Register that public debate often centers on individual systems, such as an Anduril drone or a Palantir anomaly detection system. Goodfriend said Airwars wanted to show the broader set of AI systems that can work together, or fail to work together, across battlefield surveillance, targeting and killing.

The report identifies potential failure points throughout the chain. Those include whether decision support tools are reliable, whether automated translation changes the meaning of intercepted text messages, whether risk scores built from social media data are valid, whether computer vision systems misidentify people or objects, and whether neural networks used during drone strikes degrade when GPS or electronic systems are jammed. Airwars also flags possible errors in AI-assisted battle damage assessments after a strike.

Goodfriend told The Register that the project is meant to challenge broad claims that a human remains “in the loop.” She said human review can be constrained by battlefield timing, automation bias and the fact that the material being reviewed may already have been shaped by AI systems, such as machine-translated social media posts that a reviewer cannot independently verify.

For technology companies and investors, the report is a reminder that military AI is not a clean category limited to weapons platforms. The systems described by Airwars include data processing, translation, computer vision, anomaly detection, scoring and assessment tools, many of which resemble commercial AI products when stripped of their battlefield use case.

Airwars does not claim to reveal the exact architecture of any current military deployment. Its case is narrower: the opacity around military AI makes accountability difficult, and machine-learning systems can produce errors at multiple stages before a strike occurs. Goodfriend told The Register that the point is not to treat those errors as early deployment problems that more refinement will solve, but to question whether machine learning can perform the battlefield tasks militaries expect from it.

This story draws on original reporting from The Register.