A new report, "Anatomy of an AI Kill Chain," breaks down something most people picture very differently: military AI isn't a single autonomous killer robot making a decision. It's a chain of separate AI systems -- each with its own flaws and error rate -- feeding into each other, with a human expected to catch mistakes along the way. Increasingly, that human check isn't happening the way it should.
Who's Behind the Report
The report was co-released by the AI Now Institute and Airwars, and its lead author is Dr. Heidy Khlaaf, the AI Now Institute's chief AI scientist. Khlaaf's background is notable: before working on AI safety, she worked in nuclear defense and aviation safety -- fields built entirely around rigorous, multi-layered safety review -- and later helped develop safety evaluation frameworks at OpenAI.
What the Report Found
According to Khlaaf's analysis, the individual AI algorithms used in military targeting chains have baseline accuracy rates of roughly 30 to 60 percent on their own. When these systems are chained together -- one AI's output feeding the next AI's input -- errors don't cancel out, they compound. The report's central concern isn't that AI is making unilateral kill decisions; it's "automation bias" -- the well-documented human tendency to accept a system's recommendation without independently corroborating it, especially under wartime time pressure.
Documented Examples
In Gaza, an Associated Press investigation found that Israeli Defense Forces used large language models to translate intercepted Arabic communications for targeting lists -- and documented at least one mistranslation where the word for "payment" was misread as "payload," contributing to a person being added to a target list. Separately, the U.S. Department of War has confirmed it uses Anthropic's Claude, integrated through Palantir's Maven targeting platform, in its own targeting workflows.
In Iran, a U.S. strike on a building in Minab in March 2026 killed roughly 150 people, which Iranian authorities say were schoolgirls in a girls' school at the time of the strike; the Pentagon has classified the building as a "dual-use military logistics facility." Maven is documented to use Claude to help rank targets and draft legal justifications for strikes, but neither the Pentagon, Palantir, nor Anthropic has confirmed whether AI-generated targeting was actually involved in selecting this specific building. Anthropic CEO Dario Amodei has said directly that he doesn't know what role, if any, Claude played: "We don't have access to, we don't know exactly how these models were used." Human Rights Watch has called for an independent investigation.
The Core Concern
Khlaaf's broader point isn't about any one strike -- it's structural. AI systems tend to fail in unfamiliar situations, large language models can hallucinate or fabricate information outright, and when a chain of AI tools sits between a person and a life-or-death decision, it can obscure who's actually accountable when something goes wrong. A system's designers can point to the human in the loop; the human in the loop, under pressure, often defers to the system.
Why This Matters for Families
This isn't a story about killer robots -- it's a story about how quickly "AI-assisted" decision-making can slide into "AI-trusted" decision-making once a system is embedded in a high-pressure workflow, and how hard it becomes afterward to untangle who was actually responsible. That same dynamic -- a person deferring to an AI's confident-sounding output instead of double-checking it -- shows up in far lower-stakes places too, including classrooms and homework help. Understanding how automation bias works in the highest-stakes context there is makes it easier to recognize the same pattern in far more everyday ones.