A maintainer reading a merge request can’t tell whether a person or a model wrote the diff, and nobody has to say. Debian developers voted on that through August 28, and Kurt Roeckx, the project secretary, announced the result: the winning option encourages contributors to disclose AI assistance and stops there.
Review stays where it was. A sponsor uploading a package, a security team member reading a patch for a stable release, and a downstream rebuilder pulling Debian source all work from the same signal they had a year ago, which is the diff itself.
What it requires
Not much that is new, and no endorsement either. Debian neither endorses nor prohibits generative AI, and grants that the tools can save volunteers time when used responsibly. Contributions get judged the way they always were, whatever produced them, and the person who submits stays answerable for quality, correctness, and licensing.
You read, test, and fix model output before it goes in. The adopted text puts “blindly accepting or uploading AI-generated material without appropriate human review” outside normal Debian practice, which is a polite way of saying don’t ship what you haven’t read.
Three provisions do bite. Confidential material, private communications, embargoed security bugs, cryptographic keys, and credentials stay out of third-party AI services unless someone has explicitly authorized it. Bulk work such as mass bug filing or large-scale patch submission still should seek prior discussion and a human who answers for what the automation does. And the project takes no position on whether model output is copyrightable, so DFSG compliance and existing licensing rules carry that weight unchanged.
The secrets clause
Debian’s security team works with CVE details under embargo. Pasting any of that into a hosted assistant is now explicitly out of bounds.
The industry numbers around that behavior are not reassuring. GitGuardian counted 28.65 million new hardcoded secrets in public GitHub commits during 2025, with internal repositories carrying a larger share of leaked credentials than public ones. Leaks of credentials for AI services themselves rose 81% year over year. Neither figure measures what Debian just prohibited, which is sending non-public project material to a model. They measure the adjacent habit of handling secrets carelessly inside AI-assisted workflows, and the direction is one way.
Review was already the control, and review is what compresses
OX Security analyzed more than 300 repositories, 50 of them using tools like Copilot, Cursor, or Claude. Line for line, what the models wrote was no worse than what people wrote. The trouble was volume. Reading the code, running it, and arguing about it are the steps that used to slow a team down, and they are the first ones to drop away.
That lands awkwardly on a resolution whose entire answer to AI risk is human review. Debian added no new gate. It restated the one that already exists and left contributors to apply it.
Most organizations never put AI-generated code through a review for license and IP risk. When a model reproduces something it learned from a restrictively licensed project, nothing in the output says so, and no provenance travels with it.
Debian requires licensing for everything in the archive and has now declined to say where model output falls. Those two positions have to coexist inside the head of whoever signs the upload.
