Legal scholars argue existing rules can tie wartime AI conduct to states

Unexpected AI behavior need not break the legal link between a government and its private partners. But establishing that link does not automatically prove an international-law violation. In an October 2, 2026 Lieber Institute analysis, Michael N. Schmitt and Klaudia Klonowska argue that existing state-responsibility rules are largely fit for wartime AI—even when a system behaves beyond expectations.

Two tests, not one

The authors examine private actors’ AI-related conduct during armed conflict, using the International Law Commission’s Articles on State Responsibility. This non-binding restatement provides their framework; many of its provisions are widely regarded as reflecting customary international law.

Their analysis separates two requirements. Conduct must breach an international obligation, and it must be attributable to a state—meaning international law treats it as that state’s act. Unintended AI behavior can still be attributed to a state; whether it violates international law is a separate question.

That inquiry concerns people and institutions involved in developing, deploying, operating, supervising or using a system. The authors do not treat the AI itself as a legal actor bearing international rights or obligations.

AI systems, models, and agents are not subjects of international law.

Michael N. Schmitt and Klaudia Klonowska, Lieber Institute analysis

Who controls the operation?

The authors identify Article 8 as the most likely route for attribution involving AI. It covers private actors following a state’s instructions or acting under its direction or effective control. That control must concern the specific operation, not merely a general relationship.

They illustrate the distinction with a hypothetical armed group using AI-enabled loitering munitions. If state engineers remotely set the search area, permitted targets and operating period before each mission, the authors say an attack within that operation is attributable to the state.

Merely delivering the weapons would produce a different result under that test if the group independently chose where, when and how to use them. Funding or training a group is likewise insufficient by itself. Retained override functions or kill switches could, however, indicate effective control.

Small changes, difficult judgments

A contractor’s departure from state instructions can complicate attribution under Article 8. The authors distinguish human disobedience from unexpected machine behavior: conduct clearly beyond instructions may fall outside attribution, while an incidental deviation may remain attributable.

One hypothetical involves a contractor lowering the confidence threshold that determines which AI-generated targets require human review. The mission and engagement rules remain unchanged. Yet even a small adjustment could increase false positives, making the legal significance of the change highly dependent on context.

Accountability also depends on the obligation allegedly breached. Some laws of armed conflict require intent; targeting rules usually assess decisions against information reasonably available at the time, including expected civilian harm. The authors also identify tracing conduct to its origin, especially in AI-enabled cyber operations, as a practical obstacle. Their conclusion is about the legal framework’s capacity, not a guarantee that responsibility will be established in every case.