AI on the Battlefield: How Automated Kill Chains Leave Humans Mostly Out of the Loop
Most coverage of military AI fixates on a single weapon system or a named contractor. An Anduril drone here, a Palantir targeting tool there. Airwars, the conflict transparency watchdog, thinks that framing misses the point entirely.
Their new report, Anatomy of an AI Kill Chain, maps the full sequence of decisions involved in modern military targeting and shows just how thoroughly machine learning has colonised the process. It's not one algorithm making one call. It's a stack of interdependent systems, each with its own failure modes, operating faster than any human can meaningfully intervene.
The report walks through six stages: decision support, surveillance, intelligence gathering and identification, target selection, the strike itself, and post-strike assessment. According to a recent book on US military AI cited in the report, only two of those six stages now reliably involve a human making an active decision. A third involves some oversight. The rest run on their own.
Sophia Goodfriend, a research fellow at Cambridge's Pembroke College and one of the report's co-authors, told The Register the goal was to shift the conversation away from individual products and toward the broader architecture of automated warfare. "What we wanted to do is really underscore the stack of AI systems that are upending what it means to wage war," she said.
The kill chain in the report is fictional by design. Real military systems are classified, proprietary, and largely inaccessible to journalists. But the fictitious framing is built from documented real-world components: automated translation of intercepted messages, social media risk scoring, computer vision that may misidentify people or objects, recurrent neural networks guiding drone strikes through GPS jamming, and AI-assisted damage assessments after a target has been hit.
Each stage carries its own reliability problems. Automated translations can be wrong and nobody has time to verify them. Risk scores generated from social media data carry the biases and gaps of whatever training set produced them. Computer vision systems struggle in ambiguous environments. None of these are bugs that get patched out with enough deployment. Goodfriend is explicit on this: the errors are not teething problems. They reflect inherent limitations in what machine learning can actually do under battlefield conditions.
The automation bias problem runs through all of it. When a human analyst is shown an AI-generated risk profile, under time pressure, in an unfamiliar language, the path of least resistance is to trust the output. The human-in-the-loop, in practice, often becomes a human rubber-stamping a machine's conclusion.
Military officials have previously deflected scrutiny by claiming AI's specific role in any given operation cannot be disclosed. The Airwars report is a direct challenge to that position, arguing that the public and policymakers need a clearer picture of how these systems work, what data they consume, and where they fail, before accepting assurances that meaningful human oversight is in place.
As Goodfriend put it, the private data and passive surveillance that feeds automated warfare is its "foundation and lifeblood." Making that visible is, for now, one of the few levers available to those outside the system.