How International Humanitarian Law Frames AI-Assisted Targeting Decisions

International humanitarian law (IHL) does not have a separate rulebook for AI. The core obligations that govern any attack — distinguishing combatants from civilians, weighing expected military advantage against expected civilian harm, and taking feasible precautions — apply whether a human, a targeting algorithm, or some combination of the two is involved in the decision. What changes with AI assistance is not the legal standard, but how hard it becomes to show that the standard was actually met.

The three obligations that don't move

Three IHL principles, developed long before AI-assisted targeting existed, form the backbone of how any attack decision is assessed:

• Distinction. Parties to a conflict must at all times distinguish between combatants and civilians, and between military objectives and civilian objects, directing attacks only at the former.
• Proportionality. An attack is prohibited if the expected incidental harm to civilians or civilian objects would be excessive in relation to the concrete and direct military advantage anticipated.
• Precautions in attack. Those planning or deciding on an attack must take all feasible precautions to verify targets, choose means and methods that minimize incidental harm, and cancel or suspend an attack if it becomes apparent the target is not a legitimate military objective or that the harm would be disproportionate.

These obligations sit in the body of customary and treaty IHL that predates any AI system, and they don't relax because a machine-learning model is involved somewhere in the decision chain. The question IHL asks is not "was AI used," but "was the attack lawful under these standards" — and that question has to be answered regardless of how the targeting decision was reached.

Where AI assistance actually complicates compliance

The ICRC's analysis in "Autonomous Weapon Systems and International Humanitarian Law: Selected Issues" works through where AI-assisted or autonomous targeting puts pressure on compliance in practice, rather than in principle:

• Verifying distinction gets harder to audit. A human analyst can explain, after the fact, why they judged a target to be military rather than civilian. A machine-learning classifier's judgment is often harder to trace back to a clear, reviewable rationale — which matters both for real-time verification and for any later legal accounting.
• Proportionality is a judgment call, not a calculation. Weighing "excessive" harm against "anticipated" military advantage requires contextual judgment that current systems are not designed to make on their own — a live artillery position near a school reads very differently depending on context a system may not have access to.
• Precautions assume real-time reassessment. The obligation to cancel an attack if circumstances change (a target turns out to be sheltering civilians, for instance) assumes someone is positioned to notice and act on new information up to the moment of attack — a role that gets harder to guarantee the more of the targeting cycle is automated and the shorter the window for intervention becomes.

Predictability as a legal, not just technical, requirement

This is where the IHL analysis connects directly to the human-control debate. The ICRC's broader position on autonomous weapon systems argues that a system whose effects "cannot be sufficiently understood, predicted and explained" should be prohibited outright — not as a separate new rule, but because unpredictability makes it effectively impossible for the people responsible for an attack to discharge their existing distinction, proportionality, and precaution obligations. In this framing, predictability isn't a nice-to-have engineering property; it's a precondition for a human commander being able to comply with law that already exists.

Who remains accountable

IHL attaches responsibility to people and states, not to the tools they use. A commander who deploys an AI-assisted or autonomous targeting system remains accountable for ensuring the attack complies with distinction, proportionality, and precautions — the same as if they had made the call unassisted. This is one reason the "meaningful human control" debate and the IHL-compliance debate are really the same conversation from two directions: control matters because it's the mechanism by which a legally accountable person can actually verify that these obligations were met, not an independent requirement layered on top of them.

Why this framing matters for policy debate

Treating AI-assisted targeting as a question of "does existing law apply" rather than "do we need a whole new legal regime" shapes what kind of instrument the CCW's negotiations are actually working toward. It's part of why much of the multilateral discussion — see our piece on what UN talks on LAWS have concluded so far — has focused on clarifying and reinforcing existing IHL obligations for autonomous and AI-assisted systems specifically, rather than starting from a blank page.

For the underlying vocabulary this debate uses, see what "meaningful human control" actually means, and for the technology categories involved, see autonomous vs. AI-enabled weapons. More on our editorial scope on the Agent AI Army homepage.