Human accountability for high-consequence military artificial intelligence

(a) Policy It shall be the policy of the Department that artificial intelligence may support analysis, prioritization, recommendations, automation, and other decision-support functions, but may not substitute for accountable human judgment in decisions involving the use of force, detention, domestic person-based analysis, or other high-consequence artificial intelligence applications designated pursuant to section 3. (b) Accountable human decision-Maker For each high-consequence artificial intelligence application approved for operational deployment, the Secretary shall ensure that there is a clearly identified accountable human decision-maker or accountable chain of decision-makers with authority commensurate to the operational risk presented by the application. (c) Required elements The accountable human decision-maker or chain of decision-makers under subsection (b) shall have, as appropriate to the application— (1) documented commander intent and mission parameters; (2) sufficient training on the capabilities, limitations, and failure modes of the application; (3) access to relevant information necessary to understand the basis for significant system outputs or recommendations, to the extent technically feasible; (4) authority and practical ability to pause, override, deactivate, or terminate use of the application when required by law, policy, or operational circumstances; and (5) procedures for elevating uncertainty, anomalous outputs, or significant incidents for further human review.