The Pentagon Used AI to Write the Watchdog Reports Congress Demanded
Pentagon CTO Emil Michael says an AI draft cut 200 hours of staff work to 5. The question nobody has answered is who audits the result before Congress reads it.
Pentagon CTO Emil Michael says an AI draft cut 200 hours of staff work to 5. The question nobody has answered is who audits the result before Congress reads it.
The Pentagon's top technology officer told an audience in Washington last week that artificial intelligence had done in five hours what his staff previously needed two hundred hours to produce: a report that Congress requires by law. The claim, made at a moment when the Department of Defense is restyled as the Department of War and is managing a backlog of more than 1,400 congressionally mandated reports, is being presented as a breakthrough in government efficiency. The reports AI is now drafting are the same documents that give Congress visibility into the executive branch, and nobody outside the Pentagon has audited whether those drafts meet the standard.
Emil Michael, the Pentagon's chief technology officer, made the 200-to-5-hour claim on June 12 at a Hudson Institute event in Washington covered by Ars Technica, where he described loading congressionally mandated report materials into an AI tool and receiving a draft shortly thereafter. "I have to report to Congress every year on this thing," Michael said, framing the work as a recurring obligation that his office had struggled to staff. The platform he used, GenAI.mil, the Pentagon's in-house generative-AI workspace, has been in operation since December 2025, built first on Google Cloud's Gemini for Government, a large language model adapted for federal use, and made available to members of all six military branches.
The capacity problem Michael is solving is real. The Government Accountability Office counted roughly 500 congressionally mandated reports from the Department of Defense in 2000; by 2020 the number had more than tripled to over 1,400, a workload that did not arrive with a proportional increase in legislative-affairs staff. For an office buried under that kind of statutory reporting, an AI that can produce a first draft in an afternoon is a real productivity gain, and Pentagon officials are saying so on the record.
Jacob Glassman, the deputy assistant secretary of defense for science-and-technology foundations, made a similar point at the Box Federal Summit on April 23. Glassman told the audience his team had used the new tool to produce "the best report we've written in the past five years," though he did not identify which report. The anecdote gives a second on-the-record voice to the productivity case, and the report in question is the same kind of congressionally mandated document at the heart of the AI experiment.
What neither the Hudson Institute talk nor the Pentagon press materials answer is the verification question. The reports AI is being used to draft are not internal working memos. They are the documents Congress uses to oversee the executive branch on national security, acquisitions, force structure, and a long list of other topics. When a human staffer writes the draft, a human editor typically checks the citations, the numbers, and the framing before the report goes up the chain. When an AI writes the draft, the same review may or may not happen. Michael's remarks and the Ars Technica summary of them do not describe a human-in-the-loop review process, an accuracy audit, or a disclosure practice that would let a reader know which sections of a given report were AI-generated. Glassman's "best report in five years" remark, attributed by DefenseScoop, came with the same silence on review.
That gap matters because the Pentagon's pitch to Congress is that AI is making the reporting faster without lowering the quality. The only evidence offered so far is the Pentagon's own self-assessment, including Glassman's "best report in five years" remark and Michael's 200-to-5-hour productivity claim. Neither has been independently benchmarked. The Government Accountability Office, which has tracked the growth of the congressionally mandated report backlog for two decades, has not publicly evaluated an AI-drafted report. The inspector general's office has not, on the public record, weighed in.
The question, then, is not whether the Pentagon should use AI to help draft its statutorily required paperwork. For an office buried under more than 1,400 annual reports, the productivity case is real and the technology is already deployed. The question is who reads the AI draft before Congress does, what counts as "good enough," and whether the public will be able to tell, after the fact, which paragraphs were machine-written. Until the Pentagon publishes a review standard, a disclosure practice, and a number for the actual share of reports that have moved through the AI pipeline, the 200-to-5-hour story is a productivity headline without an accountability footnote.