The Pentagon’s own responsible-AI guidelines aren’t always followed on its AI weapons projects, a Government Accountability Office report from October 14 has found. According to NPR, which covered the story on October 23, six of ten officials surveyed said their teams had “identified risk(s)” with operational use, four said requirements were not aligned with the responsible-AI tenets, and three did not know if their systems would be continuously monitored after deployment. The GAO found the oversight office running on three of its ten authorised staff. Defense Secretary Pete Hegseth, pushing an “AI-first” warfighting force, calls moving too slowly a bigger national-security risk than imperfect alignment. The Pentagon agreed to the GAO’s two recommendations but implemented neither by October 22.
What The GAO Found
The GAO interviewed ten officials directly involved in building ten AI-enabled weapons and tools — systems for automatic target identification, surveillance data processing and tactical decision-making — drawn from the Pentagon’s portfolio of more than 300 AI projects. The survey asked whether the department’s responsible-AI tenets were actually shaping development, and the answers point to a gap between policy and practice.
Six of the ten officials told the GAO their teams had “identified risk(s)” with the operational use of their systems. Four said their project requirements were not aligned with the Responsible AI tenets. And three answered that they simply did not know whether their projects would be continuously monitored once deployed — the kind of post-deployment oversight the guidelines treat as essential for AI that recommends or takes military action.
Six Tenets With Little Authority
The guidelines at the centre of the dispute date back to a 2021 Pentagon memo that set out six tenets for responsible military AI, covering areas such as human oversight, verification, real-time monitoring and cooperation with allies. The officials the GAO spoke to understood the tenets as important — but not as binding department instructions. Without formal authority, the principles function as guidance that project teams can set aside when schedules, budgets or mission pressures push the other way.
The contrast with civilian frameworks is sharp: the EU AI Act, which sets mandatory oversight rules for high-risk systems, carries legal force, while the Pentagon’s own AI weapons programme answers to principles its own builders describe as optional. That mismatch is precisely what the GAO’s survey has now put on the record.
A Skeleton Oversight Staff
The office meant to close that gap is barely staffed. The Pentagon’s Responsible AI Division was authorised ten positions in 2023, but only three were filled as of February 2026, and the department has told the GAO it is unlikely to fill the vacant positions. An oversight body operating at less than a third of its planned strength has limited capacity to review hundreds of AI projects, audit compliance or push teams toward the tenets.
Understaffing of AI governance is not only a US problem — AI laws taking shape around the world are testing how regulators staff up for the same workload on the civilian side.
Two Recommendations, Neither Implemented
The GAO issued two recommendations: clarify the responsibilities of the office overseeing responsible AI and set regular meeting schedules for it, and address the understaffing of that office. The Pentagon agreed to both — but as of October 22 had implemented neither. When NPR asked about the timeline, the Pentagon did not respond.
The non-response matters because the audit lands in the middle of an explicit push for speed. In January, Hegseth told the department that moving too slowly on AI posed a bigger national-security risk than imperfect alignment, dismissing the caution the tenets were written to protect. Last month he announced a new warfare command for autonomous weapons including drones — systems that would sit squarely inside the portfolio the GAO just flagged.
Why It Matters
The Pentagon is buying and building AI for target identification, surveillance and battlefield decisions at a scale no other military matches, and it wrote the rules meant to keep those systems under responsible control. The GAO’s findings suggest the rules are advisory in name and optional in practice, while the office enforcing them is a skeleton crew and the department’s leadership is explicitly prioritising speed over alignment. That combination — fast deployment, weak oversight, acknowledged but unresolved risks — is exactly what responsible-AI policy is supposed to prevent.
The stakes are not confined to the United States. Other militaries are racing to field comparable systems, including India’s move to add an AI layer to its defence forces, which makes the question of how AI weapons are governed a global one. When the world’s largest defence buyer treats its own safeguards as non-binding, it sets a precedent every other military will read.
Background
The Pentagon first formalised its responsible-AI principles in a 2021 memo, a move that was widely seen at the time as an attempt to set the global standard for military AI governance. Hegseth’s January 2026 AI-first memo shifted the emphasis sharply toward deployment speed, and last month’s announcement of a dedicated warfare command for autonomous weapons — drones included — signalled that the buildup is accelerating even as the GAO found oversight shrinking.
Source: NPR reporting on the GAO’s October 14, 2026 report.
0 Comments