The Pentagon Wants AI to Pick Targets. The Ethics Haven't Caught Up.

A senior Defense Department official has confirmed what many in the arms-control community have long feared: the United States military is actively exploring the use of generative AI systems to rank targets and recommend which to strike first. The disclosure, surfaced through MIT Technology Review's daily briefing, is not a rumor or a speculative scenario from a think-tank white paper. It is an acknowledgment from inside the building that AI chatbots, systems built on the same foundational architecture as the tools millions of people use to draft emails and summarize spreadsheets, are being evaluated for one of the most consequential decisions a government can make.

The timing matters. This revelation arrives alongside a separate but deeply connected story: the Pentagon's reported interest in Anthropic's Claude, one of the most prominent large language models on the market. Anthropic has publicly positioned Claude as a safety-focused model, built with what the company calls "Constitutional AI" principles designed to make the system more honest and less harmful. The tension between that branding and military targeting applications is not subtle. It is, in fact, the central contradiction that the entire defense-AI industry is currently trying to paper over.

To understand why this is significant, it helps to think about what a targeting recommendation system actually does. It ingests intelligence, sensor data, signals intercepts, and prior strike assessments, and it produces a ranked output: strike this first, then this, then this. Human operators have always done this work, and they have always done it imperfectly, under time pressure, with incomplete information, and with the full weight of legal and moral accountability sitting on their shoulders. The argument for AI assistance is straightforward: machines process faster, don't fatigue, and can synthesize more data streams simultaneously than any analyst.

But generative AI introduces a specific and underappreciated risk that older algorithmic targeting tools did not carry. Large language models are trained to produce fluent, confident-sounding outputs regardless of underlying certainty. They hallucinate. They pattern-match against training data in ways that are not fully auditable. When a model recommends a target, it cannot reliably explain the chain of reasoning that produced that recommendation in terms a military lawyer or a commanding officer can interrogate and verify. The confidence of the prose is not a proxy for the reliability of the judgment. In a courtroom, or before a congressional committee investigating a strike that killed civilians, "the model ranked it highly" is not a defense that holds.

The legal framework governing targeting, built on decades of international humanitarian law including principles of distinction, proportionality, and precaution, was designed around human decision-makers who can be held accountable. Inserting a generative AI layer into that chain does not eliminate accountability; it obscures it. And obscured accountability, in military systems, tends to migrate upward until it disappears entirely.

The Pentagon's interest in Claude specifically adds another layer of complexity. Anthropic has been more vocal than most AI companies about the risks of its own technology. Its research publications, its public communications, and its model design choices all reflect a genuine institutional concern about AI safety. Whether that concern survives contact with a defense contract is a question the company will soon have to answer publicly, if it hasn't already answered it privately.

This is not an unfamiliar dynamic in the technology industry. Google's Project Maven, the 2018 contract to apply machine learning to drone footage analysis, triggered a staff revolt and an eventual withdrawal, only for the company to quietly re-engage with defense work in subsequent years. The financial gravity of Pentagon contracts is enormous, and the national-security framing makes dissent inside companies politically costly. The pattern tends to repeat: public commitment to ethical limits, private negotiation of those limits, and eventual normalization.

What makes the current moment different is scale and capability. The AI systems being discussed for targeting in 2025 are orders of magnitude more capable, and more opaque, than the image-classification tools at the center of the Maven controversy. The decisions they are being asked to inform are not logistical. They are lethal.

The second-order consequence worth watching is not the first strike recommended by an AI system. It is the institutional habituation that follows. Once targeting workflows are built around AI recommendations, the cognitive and bureaucratic cost of overriding those recommendations rises steadily. Humans remain nominally in the loop while their practical influence over outcomes quietly diminishes. By the time that shift becomes visible, it may already be structural, and reversing structural dependencies in military bureaucracies is a project measured in decades, not budget cycles.