The same AI model that helps you draft emails reportedly helped identify and prioritize over 1,000 military targets in roughly 24 hours. According to investigative reports cited by Access Now, Palantir’s Maven Smart System — integrated with Anthropic’s Claude — served as a “reasoning mechanism” during U.S. operations against Iran. That dual-use reality is why more than 200 civil-society organizations, led by Access Now and Amnesty International, are now demanding governments and tech companies immediately stop deploying AI in the military kill chain.

“Death Conveyor”: How AI Compressed the Kill Chain

Targeting algorithms have collapsed military decision cycles from days to minutes — and the human judgment that once filled that gap is disappearing with them.

The kill chain is military shorthand for a grimly simple sequence: find a target, decide it matters, strike it. AI has collapsed that timeline from days to minutes — sometimes seconds, according to the Center for International Policy. Reports describe U.S. and allied forces striking close to 2,000 targets in four days during recent hostilities. This isn’t a software upgrade. It’s a conveyor belt where life-and-death decisions move at speeds that make genuine human deliberation physically impossible.

The coalition’s demands are specific:

• An immediate halt to AI in military targeting, including target generation and prioritization
• Full transparency from states on current AI use in active hostilities
• Binding obligations on tech companies to prevent their AI from contributing to violations of international humanitarian law
• An end to what Access Now calls “hollow ethical declarations” from companies simultaneously profiting from military contracts

“AI has become the new gunpowder, and tech companies are the new arms dealers.” — Marwa Fatafta, MENA Policy and Advocacy Director, Access Now.

The legal gap is significant. No binding international treaty currently governs AI in military targeting. International humanitarian law requires distinguishing civilians from combatants and ensuring proportional force. Those are context-dependent judgments that algorithms cannot reliably make. SIPRI research shows algorithmic bias can skew target identification. The ICRC warns these risks are “permanent” and cannot be engineered away. Commanders who deploy systems with known targeting defects could face war-crimes liability under existing command responsibility doctrine.

Meanwhile, according to reporting cited by Access Now and others, OpenAI, Anthropic, and Meta have each entered partnerships placing their models into defense workflows. The BBC reported that the U.S. administration designated Anthropic a “supply chain risk” over its resistance to unrestricted military access, subsequently ordering federal agencies to phase out its tools. The responsible-AI branding and the defense contracts are now on a direct collision course.

What “Meaningful Human Control” Actually Means

The debate isn’t whether humans are technically present in targeting decisions — it’s whether they’re meaningfully making them.

Militaries insist humans stay “in the loop.” But when an operator rubber-stamps AI-generated target lists at machine speed, that loop is a fiction — like calling autoplay an active viewing choice. The Center for International Policy recommends hard distinctions: preauthorized defensive systems can operate within tight automated boundaries, but offensive strikes should require explicit human approval. High-consequence operations must retain what the Center calls “a human face and a human name at the end of the chain of action.”

The engineers who once organized against Project Maven are watching their successors ship models directly into kill chains. If the tech industry cannot hold the line between a productivity tool and a targeting engine, regulation will arrive — potentially through a binding UN treaty process that advocates are already pushing. The only question is whether it arrives before the next thousand targets are queued.