The Pentagon now has 1.5 million of its 3.5 million staff using AI tools every single day. That is nearly half the entire department, up from just 80,000 users when the platform launched six months ago. The jump happened after the military made Google's Gemini available across its standard, non-classified computer networks and clarified what staff were allowed to use it for.
The platform, called GenAI.mil, gives everyone from soldiers to civilian contractors access to the same AI tools that businesses use: drafting documents, summarising meetings, building budgets. Routine office work. The Pentagon's chief technology officer, Emil Michael, said staff are also using it for job descriptions and other administrative tasks.
The headline use case he chose to publicise, however, is more significant. He described feeding source documents into the AI and having it produce the official written reports the Department of Defense must deliver to Congress by law each year. He put the time saving at 200 hours of staff work reduced to five.
For context: those reports are not optional paperwork. They are the formal mechanism by which Congress keeps watch over the military. When Congress funds a weapons programme, a new policy, or a military operation, it often requires the Pentagon to report back on what happened. If those reports are AI-generated and not carefully checked, the oversight chain has a new weak link.
The Pentagon says its officials review AI output before it goes out. That review step is everything. AI tools, including the ones on GenAI.mil, can produce text that sounds authoritative while containing details that are subtly wrong or simply made up. This is a known and documented problem with all current AI writing tools, not specific to any one product. A report that looks complete and well-structured can still contain figures, references, or summaries that a rushed reviewer would not catch.
There is a broader pattern here worth watching. The speed gain is real: cutting 200 hours to five is not a rounding error. For a department with hundreds of mandatory reports a year, that is a meaningful reduction in staff time. The question is whether the quality controls are scaling at the same pace as the adoption numbers.
For business operators in any regulated industry, this is the template playing out in public. Governments are adopting AI writing tools for official documents faster than oversight frameworks can catch up. The House Armed Services Committee is already considering requiring the Pentagon to build a formal incident-reporting system for AI failures. That suggests even legislators see the gap.
The practical signal for any organisation that reads, responds to, or relies on government-produced documents: the documents themselves may now carry a new category of error, one that looks like confident, well-formatted writing but was never verified against primary sources. That is a different kind of risk than a typo. It requires a different kind of reading.