How much should military AI agents decide for themselves? UNIDIR recommends that states explicitly define authorization limits and intervention thresholds. Its national AI strategy guide, published on October 19, also calls for rollback mechanisms and conditions governing deployment.
Authors Yasmin Afina, Jan Hendrik Mannsperger and Edward Madziwa give multi-step agents a separate section. Machine-speed actions, they argue, can challenge assumptions about human control, command structures and responsibility.
Another section warns about excessive reliance on automated decisions and responsibility becoming diffuse. The recommendations ask states to account for the speed and scale of the systems they oversee.
The 81-page edition revises earlier drafts following consultations and a literature review. It presents the authors’ recommendations; the publication does not establish whether states have implemented them or how effective the proposed limits would be.