AI in the Kill Chain
The integration of artificial intelligence into military targeting has been one of the most closely watched and least transparent developments in modern warfare. Now, for the first time, a Defense Department official speaking on background has provided specific insight into how generative AI — the same technology that underlies ChatGPT and other large language models — is being used to assist targeting decisions in the ongoing U.S. military campaign against Iran.
According to the official, who requested anonymity due to the sensitivity of the topic, generative AI systems are being used as a conversational analysis layer on top of existing intelligence and targeting data. A list of potential targets is fed into the AI system, which is asked to analyze the information, prioritize targets, and generate recommendations based on factors including aircraft positioning, mission objectives, and available intelligence. Human operators then review, evaluate, and take responsibility for acting on these recommendations.
Two Different AI Technologies
The official's account highlights an important distinction that is often lost in public discussions of AI in warfare: the Pentagon is deploying two fundamentally different types of AI for related but distinct functions.
Project Maven, operational since at least 2017, uses computer vision and machine learning to process the enormous volumes of imagery and sensor data collected by surveillance systems — drone footage, satellite imagery, signals intelligence. Maven identifies potential targets within this data and presents them to human operators through a map-based interface. This is AI as data processor and pattern recognizer, operating on well-defined, supervised tasks.
Generative AI — systems built on large language models — is different in kind. These systems are conversational, flexible, and capable of reasoning across diverse types of information. They can synthesize intelligence from multiple sources, generate written assessments, and answer open-ended questions. But they are also less battle-tested, less transparent in their reasoning, and more prone to the kinds of confident-but-wrong outputs that experts call hallucinations.
The Human-in-the-Loop Question
The official was explicit that humans remain responsible for evaluating AI recommendations and making final targeting decisions. This human-in-the-loop framing is standard in Pentagon communications about AI, and it reflects both genuine policy commitment and operational reality: no military commander is going to outsource life-and-death decisions entirely to an algorithm.
But the practical question of how much deference humans give to AI recommendations — especially under time pressure and cognitive load — is more complex than any official statement can capture. Research on decision-making under uncertainty consistently finds that when authoritative-seeming systems provide confident recommendations, human oversight tends to become cursory rather than genuine.
Claude, OpenAI, and the Race for Pentagon Contracts
The official's comments come amid a dramatic reshaping of commercial AI's relationship with the U.S. military. Anthropic's Claude was the first large language model approved for classified Pentagon use and was reportedly deployed in operations in Iran and Venezuela. But following a dispute with the Pentagon over whether Anthropic could limit how its model was used, the Defense Department designated Anthropic a supply chain risk, and President Trump called for government use of the company's AI to end within six months.
OpenAI filled the gap, announcing an agreement on February 28 for the military to use its technologies in classified settings. Elon Musk's xAI has reached a similar deal for Grok. The speed of these commercial relationships — and the limited public scrutiny they have received — has drawn criticism from AI safety researchers and arms control experts who argue that AI integration into lethal military systems is proceeding faster than governance frameworks can accommodate.
This article is based on reporting by MIT Technology Review. Read the original article.
