Governments have agreed to spend 2027 and 2028 drafting rules for lethal autonomous weapons, after a Geneva conference ended without an immediate mandate to negotiate a legally binding treaty. The unresolved question is what legal force those rules will have.

Switzerland's Foreign Ministry, whose ambassador chaired the conference, confirmed the outcome on November 21. It said the 128 parties to the Convention on Certain Conventional Weapons agreed to develop an instrument regulating these weapons, with its legal form to be decided after the drafting work.

What the Geneva decision changes

The facade and lawns of the Palais des Nations in Geneva
An archive view of the Palais des Nations, first uploaded to Wikimedia Commons in 2005.Photo: © Yann Forget / Wikimedia Commons, CC BY-SA 3.0

The conference moved beyond the question in our earlier preview of the talks: whether governments would authorize further drafting. They now have a two-year assignment, but the agreement leaves open whether the result will bind participating countries.

Decisions required consensus. Switzerland said a majority wanted an immediate mandate for a legally binding instrument, but that proposal could not secure agreement. Swiss and Peruvian diplomats worked on the compromise through the early hours of Saturday, November 21.

What makes a weapon autonomous

A white Phalanx close-in weapon system mounted aboard USS Denver against a blue sky
A Phalanx close-in weapon system aboard USS Denver on November 14, 2007. This archive photograph does not show a newly announced deployment.Photo: Crassic / Wikimedia Commons, CC BY-SA 3.0

The International Committee of the Red Cross defines these systems by their targeting functions. Once activated, they can select a target and use force without a person choosing that specific target at the moment of attack.

A remotely piloted armed drone whose operator identifies the target and decides when to fire falls outside that definition. A drone that flies itself does not necessarily make its own targeting decisions, either. Navigation and the decision to attack are separate functions.

Autonomous weapons also need not use machine learning. The ICRC identifies existing examples such as missile-defense systems, tank protection systems and certain sensor-fused munitions. Many rely on rules and operate against a narrow range of military targets.

The dispute therefore reaches beyond a particular AI model. It concerns which targeting decisions machines may make, where they may operate and how people can constrain or interrupt their actions.

How U.S. policy compares

The Pentagon's January 2023 directive update called for autonomous and semiautonomous weapons to let commanders and operators exercise appropriate human judgment over force. It also required realistic performance testing and care consistent with the law of war, treaties, safety rules and rules of engagement.

An October 5 Congressional Research Service report explains the distinction between judgment and direct control. Human involvement can include deciding how, when, where and why a weapon is deployed. That requirement does not mean a person must approve each individual engagement.

The report distinguishes fully autonomous systems from human-supervised systems, where an operator can monitor and stop an engagement, and semiautonomous weapons, which engage targets or groups that a person selected.

Congress has separate oversight requirements. The CRS report says lawmakers must receive notice of changes to the directive within 30 days, and annual reports on U.S. approval and deployment of lethal autonomous weapons through December 31, 2029.

Those U.S. procedures and the proposed international rules address different obligations. A domestic approval process sets requirements for American military systems; an international instrument would establish commitments among participating governments.

What the Red Cross wants restricted

In its November 21 response, the ICRC criticized the failure to secure a clear, immediate mandate for binding negotiations. It wants a treaty that prohibits unpredictable autonomous weapons and those designed to target people directly, while restricting other autonomous weapons.

The ICRC says deciding whether a person is surrendering, wounded or directly participating in hostilities requires judgments about conduct and context. Those circumstances can change between activation and an attack. It argues that predicting a weapon's behavior becomes harder as systems grow more complex and operate in populated environments.

For other autonomous weapons, the ICRC's 2026 advocacy paper proposes limits on the types of targets and the geographic scope, duration, situations and scale of use to keep people in control.