The Ethics of Autonomous Weapons Systems

What It Covers

Hebrew University law professor Yuval Shani joins host Matt Merrill to examine how AI is transforming warfare faster than international law can regulate it. They cover existing autonomous weapon systems like Israel's Harpy drone, the US military's JADC2 program, decision-support AI already used in active conflicts, and the accountability gaps these technologies create under international humanitarian law.

Key Questions Answered

• •Autonomous weapons threshold: Fully autonomous humanoid battlefield robots remain theoretical, but partial autonomy already exists operationally. Israel's Harpy drone loiters, identifies radar signals, and self-destructs without human input. The US military's CODE program develops drones for communication-denied environments, and JADC2 aims to connect threat identification directly to weapons deployment — all without a human pulling the trigger at the critical moment.
• •Decision-support AI as the real frontier: The more immediate legal and ethical challenge is not killer robots but AI recommendation systems already deployed in active conflicts. The US military reportedly uses Anthropic's Claude for targeting. Ukraine uses GIS ARTA, described as "Uber for artillery," to route fire decisions algorithmically. These systems operate now without a clear legal framework governing how much weight human operators must give their outputs.
• •The floor-ceiling collapse problem: International humanitarian law sets a minimum legal floor — do not deliberately target civilians, apply proportionality analysis — but historically soldiers exercised restraint beyond that floor due to empathy, uncertainty, or de-escalation incentives. Autonomous systems eliminate that discretionary layer, effectively collapsing floor and ceiling into one, transforming war from a human endeavor into what Shani calls industrial-scale machine-directed killing.
• •Speed outpaces oversight: The Israeli military publicly stated that AI reduced target generation from 100 targets per year to 100 targets per week. In active combat, intelligence officers reportedly had as little as 20 seconds to approve or reject AI-generated target recommendations before passing them to weapons teams. At that decision velocity, meaningful human review becomes procedurally nominal rather than substantively real, regardless of what policy mandates.
• •The accountability gap incentivizes AI use: When AI mediates a war crime, prosecuting anyone becomes nearly impossible. Criminal liability under laws of war requires proving intent — that a person knew their action would likely cause a specific outcome. Because AI systems are black boxes that operators do not fully understand, that intent threshold cannot be met. This accountability vacuum paradoxically creates a legal incentive to route lethal decisions through AI rather than away from it.
• •Proliferation exceeds nuclear containment: Unlike nuclear weapons, which require state-level infrastructure and remain limited to roughly eight to ten countries under nonproliferation frameworks, military AI is built on dual-use commercial models already publicly accessible. Claude, a consumer product, is already used for targeting. Combined with commercially available drones, any actor can assemble a functional lethal autonomous system at low cost, making containment through international treaty structurally unenforceable.