The Pentagon's guidelines for "responsible AI" aren't always followed, report shows

A federal watchdog has found that the Pentagon's own ethical rules for artificial intelligence are applied unevenly to the weapons and tools now under development. The Government Accountability Office surveyed officials on 10 of more than 300 AI projects, including systems that spot targets automatically. Six officials said their teams had flagged risks of fielding the technology in real operations. Four said project requirements did not match the department's responsible-AI principles, and three did not know whether their systems would be monitored after deployment, according to NPR.

The principles come from a 2021 memo setting out six tenets. Alissa Czyz, who led the review, noted the memo is guidance, not a formal instruction. The division that helps train staff on the guidelines was cleared in 2023 to hire 10 people but had filled only three posts as of February, and further hiring is unlikely, GAO says.

GAO urged the Pentagon to define the oversight office's duties and meeting schedule and to fix the staffing gap. The Pentagon agreed but had made no changes by Oct. 22. Defense Secretary Pete Hegseth has argued the risk of moving too slowly outweighs imperfect alignment.

Summarized by PolitiKool from NPR’s reporting. Read the full story at NPR →