Statement by Stop Killer Robots to Session 4 of the informal exchanges on AI in the military domain

The following statement was submitted by Stop Killer Robots for ‘Session 4: The Life Cycle of AI in the Military Domain (Part 1)’ of the informal exchanges on AI in the military domain, 15-17 June 2026

Thank you, Chair.

We would like to take the opportunity of this session to raise concerns related to the data that is used in AI systems for designation and categorisation in targeting, which can be problematic in two ways.

First, the processes by which data is obtained can raise numerous concerns.

Data sources are often opaque. Reporting and available information suggests that the data used in various AI systems: may not be appropriate to the use case; can be built on a variety of assumptions which may not apply to all cases for which an automated system is to be used; or can be outdated or otherwise incorrect. All of these factors can increase the risk of errors, inaccuracy or indiscriminate effects.

Data might also be gathered through mass surveillance inconsistently with human rights law, or repurposed from commercial or public applications in the civilian sphere with little awareness by rights holders and potentially obtained without their informed consent.

Though some of these problems apply to processes that do not use AI and automation, they can be amplified by them.

In addition, available data often does not adequately represent the factual and complex human contexts on the ground in armed conflict, and all data inevitably reproduces the biases of its context, including sexism, racism, ableism, and other forms of discrimination. Human decision-makers in the ‘kill chain’ may have neither sufficient knowledge of the provenance of a system’s training data, nor the ability to verify outputs and take sufficient time to form an independent understanding of the context to make a meaningful decision.

This must be prevented.

Using opaque systems that people, notably system operators, don’t functionally understand is unacceptable in the use of force. Systems in the use of force must always be predictable, reliable, explainable and traceable: however, this is technically impossible with current forms of probabilistic AI – providing a clear red line that must be drawn.

Thank you, Chair.