Defense ministries drafting policies to ban AI-assisted software development from procurement are writing rules for a world that no longer exists. Microsoft's CEO told an audience in April 2025 that 20 to 30 percent of code in some company repositories is now AI-generated - and there is no reliable method to measure it after the fact, "Hvylya" reports, citing a War on the Rocks analysis by Markus Sandelin, AI Lead at the NATO Communications and Information Agency.
The problem runs deeper than any single vendor. The most popular AI coding assistants count tens of millions of users and generate code at a pace no manual review can match. The open source foundations underneath defense systems - including Linux itself - are increasingly maintained by contributors using these tools. By the time any of this reaches a defense application, the code has passed through dozens of AI-touched links in a supply chain no nation currently tracks.
"Code is code. There is no watermark. There is no signature," Sandelin wrote. A function written by a human and the same function generated by an AI are identical artifacts. Provenance tracking only works if recorded at the moment of generation, in a chain of custody that no party in the supply chain currently maintains.
The ban creates a second problem that may be harder than the supply chain gap. Developers use the tools regardless of organizational policy because the productivity gain is too large to leave on the table. A developer ignoring the ban ships features faster, closes tickets sooner, and gets praised for output. The ban does not prevent use - it prevents visibility and pushes the tools underground, eliminating logging, review, and institutional awareness.
Defense ministries should redirect the energy currently spent on unenforceable prohibition toward verification infrastructure, Sandelin argued: using multiple AI models against each other for review, demanding tool-level provenance records from suppliers, and investing in runtime monitoring that watches how code actually behaves in production. "The honest standard is 'better than nothing,'" he wrote, "and the gap between nothing and these measures is the difference between structural indefensibility and a reasonable chance of detecting compromise."
Also read: why the Pentagon's AI targeting systems lack training, doctrine, and safety rules.