New Report Warns AI 'Kill Chains' Are Driving Civilian Deaths in Gaza and Beyond

Militaries are feeding decisions about who lives and who dies through chains of AI systems that were never designed to work together, and each link in that chain is unreliable on its own.

That is the core finding of "Anatomy of an AI Kill Chain," a joint report from the AI Now Institute and the conflict-monitoring group Airwars. Heidy Khlaaf, chief AI scientist at AI Now, laid out the findings in an interview with Democracy Now!

Not One Killer Robot, But a Chain of Them

Khlaaf said the public imagines AI warfare as a single autonomous weapon making a kill decision. The reality, she said, is a pipeline: multiple algorithms, from older computer vision systems to newer large language models, feeding outputs into each other before a human ever sees a recommendation.

Each algorithm in that chain, according to Khlaaf, carries its own error rate and its own blind spots. When they are stacked, the flaws compound. The report argues this stacking effect, not any single rogue system, is what has repeatedly translated into civilian casualties in modern conflicts.

Khlaaf noted that militaries have used AI in some form since the 1960s. What has changed, she said, is the push to fold generative AI tools like ChatGPT-style large language models directly into targeting and decision-making, a shift she and her co-authors describe as a new phase of an "AI arms race."

The Illusion of a Human 'In the Loop'

U.S. officials have repeatedly insisted a human remains the final check. Defense Secretary Pete Hegseth has said the Pentagon intends to become an "AI-first warfighting force" from its back offices to "the tactical edge on the frontlines." CENTCOM Commander Admiral Brad Cooper has said advanced AI tools compress what once took hours or days into seconds, while stressing that humans "will always make final decisions on what to shoot and what not to shoot."

Khlaaf's research complicates that reassurance. She pointed to well-documented "automation bias," the tendency of human operators to accept an algorithm's recommendation without independently corroborating it. If a system is designed to move at machine speed and a human is expected to sign off in seconds, she argued, the distinction between a human-supervised "decision support system" and a fully autonomous weapon becomes, in her words, superficial in practice.

Khlaaf also flagged that Russian and Chinese information operations have attempted to seed propaganda designed to skew the outputs of large language models, a warning that speaks to how fragile the data feeding these systems already is before it ever reaches a battlefield.

Why It Matters for Gaza

The report's title and framing situate this critique squarely within conflicts including Gaza, where Israeli military targeting has drawn scrutiny for its scale and pace. The mechanism Khlaaf describes, low-reliability algorithms chained together and rubber-stamped under time pressure, is precisely the accountability gap human rights monitors have pointed to when questioning how targeting decisions in Gaza are made and reviewed.

The report does not itself supply new casualty figures. Its contribution, according to Khlaaf, is structural: showing regulators, journalists and the public how an AI-assisted kill chain actually functions, so that when civilians die, the question of which system, and which human, failed is no longer unanswerable.