The US Military Is on an AI Sprint. Its Accountability Mechanisms Are Still in the Starting Blocks.

The Pentagon has built more than 100,000 user-created AI agents in the five weeks since its Agent Designer tool became available, Deputy Assistant Secretary of Defense Jacob Glassman said, while the department's accountability mechanisms have not kept pace with adoption. Glassman attributed the surge to workforce attrition from the Deferred Resignation Program and operational urgency during active military operations, describing the Pentagon's approach to GenAI.mil as "cattle-driving, as we say, everything in GenAI.mil." The episode, in which a senior Pentagon official directed staff to use GenAI.mil to meet a congressional reporting deadline without a described review process or accuracy metric, illustrates a governance bypass, the article argues.

At a federal technology summit https://www.defenseone.com/defense-systems/2026/04/pentagon-adds-googles-latest-model-genaimil-usage-soars/413126/ earlier this year, a senior Pentagon official described a moment his office was proud of. Facing a congressional reporting deadline, and with a workforce hollowed by the Deferred Resignation Program, he directed his staff to use GenAI.mil to meet the deadline. A week later, the staff returned. Not only did the staff members generate the report, but he recalled them saying that it was the best report they had written in five years. He called it incredible, but as he described it there was no mention of a review process or any accuracy metric. He went on to note that in the five weeks since the Agent Designer tool became available, Department of Defense personnel had built more than one hundred thousand user-created AI agents https://breakingdefense.com/2026/04/pentagon-workers-vibe-code-100000-ai-agents-to-use-on-unclassified-networks/ . The Pentagon’s approach to adoption, he told a reporter https://defensescoop.com/2026/04/23/pentagon-uses-genai-mil-to-create-agents/ on the sidelines, was straightforward: “cattle-driving, as we say, everything in GenAI.mil.” The celebration was genuine. However, the episode reveals the problem: The speed of AI advancement and Pentagon adoption has vastly outpaced the development of accountability mechanisms that should accompany it. Speed and governance are productive when designed together. The legacy procurement system was not accountable. It was compliant. And those are not the same thing. But compliance was never empty ritual. The forms, the sequential sign-offs, the paper trail—that apparatus was how the Department of Defense built checks and balances into the fabric of routine work for the better part of a century. Process-heavy acquisition produced diffuse responsibility dressed as rigor, and senior leaders are right to dismantle a system that had grown slower than the threats it was meant to manage. But dismantling the bureaucracy is not the same as dismantling what the bureaucracy did. The current sprint has built speed. It has not built anything to carry the checks the old system used to carry by default. Three Pressures, One Bypass The GenAI.mil rollout unfolded under pressures that would individually stress any governance framework. Together they constitute a governance bypass. The first is workforce attrition. Deputy Assistant Secretary of Defense Jacob Glassman named it plainly at the Box Federal Summit in April https://defensescoop.com/2026/04/23/pentagon-uses-genai-mil-to-create-agents/ . The Deferred Resignation Program hollowed out capacity, and along with it, judgment. Formal governance frameworks document processes and procedures; they do not capture the institutional knowledge experienced practitioners carry, such as knowing which outputs to distrust, which approvals to route, and which congressional sensitivities to observe. That tacit knowledge layer departed before the AI arrived. No requirement existed to transfer governance functions explicitly before positions were eliminated. The AI was positioned as the solution to the capacity problem, and the governance gap it inherited went unaddressed. The second pressure is operational urgency. Glassman framed the agent milestone against active military operations https://defensescoop.com/2026/04/23/pentagon-uses-genai-mil-to-create-agents/ , noting the department was operating in a constrained environment while at war. Military organizations are designed to subordinate process to mission when necessary, and that design feature functions correctly in kinetic operations. The challenge emerges when urgency logic migrates into information governance. The congressional report case is exactly this pattern: a mandatory deliverable, a depleted staff, an available tool, and a senior leader’s authorization to proceed. No one made the wrong decision according to their immediate professional logic. The governance failure was structural, and the operational context made it invisible. The third pressure is mandate compliance. Since December 2025, the Department of Defense has pursued a saturation campaign for AI adoption. Secretary Pete Hegseth declared the department was pushing all its chips in. Every military service except the Coast Guard designated GenAI.mil https://defensescoop.com/2026/02/02/military-branches-genai-mil-enterprise-ai-adoption/ as its preferred enterprise platform. The platform has since grown past 1.7 million users https://shattered.io/pentagon-chatgpt-grok-genai-mil-2026/ , roughly half the department’s workforce, and in August expanded from a single model to three, adding ChatGPT and Grok alongside Gemini. When the senior-most level of an institution signals that adoption is the performance metric, the practitioners closest to implementation problems carry the least organizational power to slow the pace. A Known Trap in New Clothes None of this is unprecedented. More than a decade ago, Leonard Wong and Stephen Gerras https://press.armywarcollege.edu/monographs/466/ documented what happens when institutional compliance demands consistently exceed institutional capacity. Leaders, faced with requirements they cannot fully satisfy, develop what the authors called ethical fading: a gradual decoupling of compliance reporting from actual compliance, encouraged and sanctioned by the institution itself. The facade, as one staff officer told the researchers, goes all the way up. Congressional queries are answered on time. Readiness reports brief green. The institution keeps rolling. The GenAI.mil case shows how the Wong-Gerras dynamic could acquire a unique amplifier. The staff did not produce a dishonest report in the traditional sense. They produced an AI-assisted product that looked better than what an exhausted, understaffed team would have generated manually, was delivered faster, and was received as evidence that AI is working. The feedback loop ran in exactly the wrong direction. Scrutiny should have followed the celebration; instead, the next adoption milestone followed. The critical difference from Wong and Gerras matters here. In their model, practitioners know they are equivocating. They feel the moral weight of the compromise even while making it. In the AI-amplified version, that awareness may not be present at all. The practitioner may genuinely believe the output is high quality, and the official celebrating the report clearly did. Imperceptible ethical fading is a more dangerous condition than the kind that produces discomfort. Discomfort creates the possibility of correction; satisfaction precludes it. The Structural Conditions The three pressures under which the department’s accelerated AI adoption played out do not inevitably produce a bypass. They require preexisting structural vulnerabilities to exploit. The first is fragmented authority. No single entity owns the AI governance space. Oversight is distributed https://defensescoop.com/2025/08/15/feinberg-cdao-realignment-shakeup-dod-ai-enterprise/ across the Chief Digital Office and Artificial Intelligence Office, service chief information officers, combatant commands, and individual unit commanders, each holding partial jurisdiction and none holding full accountability. The second is accountability and attribution ambiguity, illustrated precisely by the congressional report case. Who is accountable for that product? The deputy assistant secretary who directed its production? The staff who constructed the prompt? The office that deployed the platform? The vendor whose model generated the content? Every actor in that chain holds a defensible argument that accountability belongs elsewhere. This is organized irresponsibility in its operational form: Everyone is nominally responsible; therefore no one is actually accountable. AI adoption at scale institutionalizes the condition. The FY2026 National Defense Authorization Act mandates governance frameworks through Sections 1512 and 1533 https://www.akingump.com/en/insights/alerts/congress-moves-forward-with-ai-measures-in-key-defense-legislation , but the timelines defer the reckoning: DoD owed Congress its AI cybersecurity review by August 2026, yet the Section 1533 assessment framework is not due until January 2028. The problem is operating now. The third vulnerability is agent proliferation without registration. The 103,000-agent figure https://breakingdefense.com/2026/04/pentagon-workers-vibe-code-100000-ai-agents-to-use-on-unclassified-networks/ cited in April was a five-week snapshot; subsequent reporting describes hundreds of thousands of agents https://shattered.io/pentagon-chatgpt-grok-genai-mil-2026/ deployed across the platform. No centralized registry, audit trail, or named accountable official per agent has been announced. In personnel terms, this means organizational actions are increasingly performed through AI agents without any corresponding mechanism for attributing ownership. Command responsibility ordinarily depends on identifiable delegation and attributable organizational action; the architecture described here makes that attribution difficult. The organizational logic used to attribute responsibility for human activity has not been extended to AI-enabled activity conducted in the institution’s name. The vendor layer deserves a specific note, because it demonstrates that the gap is a choice rather than an inevitability. Three commercial models now operate on the platform under agreements permitting use for any lawful government purpose https://www.computing.co.uk/news/2026/government/google-signs-ai-deal-pentagon , with end-use accountability belonging explicitly to the department. But those terms proved negotiable. In March, following public pressure, OpenAI amended its agreement https://www.notebookcheck.net/OpenAI-adds-new-guardrails-to-Pentagon-deal-after-US-surveillance-backlash.1240281.0.html to prohibit its ChatGPT product to be used for domestic surveillance of US persons and to require a separate agreement for intelligence agency use. Accountability boundaries were achievable when a vendor insisted on them. The relevant question is not whether commercial AI can be bound, but whether the department requires it. These structural conditions compound in real operational decisions with results that cannot be undone. Four Asks That Do Not Require an Act of Congress Practitioners at every level hold authority to act inside existing structures without waiting for framework completion. First, establish an AI provenance standard for congressional and executive products. This is a professional responsibility norm, not a technical requirement. Any AI-assisted product delivered to an oversight principal, whether Congress, the inspector general, or a combatant command, should identify the model used, the prompting authority, and the human review step. This is the citation standard applied to a new form of knowledge production, and the profession already norms attribution. The operational side implemented it more than a year ago: The National Geospatial-Intelligence Agency adopted a disclosure template https://www.executivegov.com/articles/nga-standardized-disclosure-ai-generated-geoint for machine-generated intelligence products so that combatant commanders, the secretary of defense, and the president would understand the risk continuum they were operating under. No equivalent standard exists for AI-generated products delivered to Congress. Second, require governance transfer before workforce reductions take effect. Before any position is eliminated in a role exercising judgment over information quality, outputs, or compliance, the governance functions that person performed must be explicitly reassigned to a named successor. The Deferred Resignation Program eliminated the addresses without transferring the mail. AI does not inherit judgment; someone must exercise it. Third, maintain a command-level agent registry. Commanders hold accountability for what happens in their formations, and that accountability cannot be exercised over agents they cannot identify. A registry recording agent name, creating user, stated purpose, activation date, and status is a minimal administrative control. Hundreds of thousands of unregistered agents create a potentially enormous volume of AI-enabled activity for which commanders lack a standardized record of ownership, purpose, and status. Fourth, pair every adoption metric with an accountability metric. Users, sessions, and agents built are throughput measures. They indicate how fast the herd is moving and say nothing about direction or destination. For every adoption milestone, require a corresponding accountability milestone: a named reviewer, an error correction pathway, an accountable official. The congressional report was celebrated for its quality with no accountability metric attached to it and no disclosed baseline against which that judgment could be evaluated. The distinction that matters is not whether a human is involved. It is whether the human exercises judgment before the output acts or after. The Mission Command Problem Secretary Hegseth’s January 2026 AI strategy https://media.defense.gov/2026/Jan/12/2003855671/-1/-1/0/ARTIFICIAL-INTELLIGENCE-STRATEGY-FOR-THE-DEPARTMENT-OF-WAR.PDF calls for accountable leaders who move fast, remove blockers, and deliver measurable outcomes. That is mission command logic applied to AI adoption, and it is precisely right as a principle. Mission command accepts decentralized execution without abandoning attributable responsibility. Commanders need not know every action their subordinate elements take, but they must be able to establish intent, assign responsibility, and assess whether actions taken in their formations remain consistent with that intent. With hundreds of thousands of agents with no registry, provenance standard, or governance transfer requirement, commanders cannot reliably identify the risks they are underwriting. The mission command logic the administration correctly applies to warfighting does not transfer automatically to AI-enabled staff work. Someone must deliberately extend it. The question is not whether the decision loop closes. It will. The question is whether the governance apparatus is inside the loop when it does, and whether the profession normalizes its absence before someone with standing asks where the accountability went. The herd is moving. Mission command requires someone to know where. Right now, no one does. Colonel Josh Goodrich is a Massachusetts Army National Guard engineer officer with a PhD in management from the University of Massachusetts Lowell, Manning School of Business. His research focuses on organizational governance, authority, and accountability in complex ecosystems, including the national security enterprise. The views expressed are those of the author and do not reflect the official position of the United States Military Academy, Department of the Army, or Department of Defense. Image credit: Jim Shea, US Army