Debian voted on AI, and the winning option discourages telling anyone you used it

Debian's General Resolution closed August 29. Option 5, "Responsible Use of Generative AI," won the Condorcet ballot with 281 votes out of a 1,045-developer electorate, and all eight options cleared the 48.49 quorum. That part is procedural. The text is what caught me.

The winning policy holds contributors to identical quality, correctness and legal standards regardless of what tool produced the patch, and puts responsibility squarely on the submitter. Then it discourages disclosing that you used AI. Not "doesn't require." Discourages. The Linux kernel went the other way and requires attributing which model was used. Two of the largest projects in open source read the same problem and reached opposite conclusions about whether provenance is information or noise. (Debian)

The two stricter anti-AI options, 1 and 3, both lost to None of the Above. That's the result I'd have gotten wrong if you'd asked me to predict it a month ago. The HN thread ran 490 points and 456 comments, so the disagreement is real even where the ballot wasn't close.

Set that against a survey published a day earlier. Someone reviewed 120 open source projects and found 37 with total AI contribution bans, naming GCC, QEMU, SDL, Gentoo, Zig and Ghostty, plus platform-level bans at Codeberg and Sourcehut and a Flathub policy that covers code, docs, bug reports and review comments. Greg Kroah-Hartman is quoted saying current tools produce results that are at least a third flat-out wrong or harmful. The stated driver in nearly every case is maintainer review time consumed by low-quality submissions. (optimizedbyotto.com)

So roughly 31% of surveyed projects ban it outright, and the largest distribution in Linux just voted that you shouldn't even mention it. If you contribute AI-assisted patches, you now need a per-project lookup table, and the answers range from "attribute the model by name" to "don't mention it" to "we will close your PR on sight."

My read on why Debian went where it did: disclosure creates a two-tier review queue. Once a patch is labeled AI-assisted, maintainers triage it differently regardless of quality, and the label becomes a proxy nobody validated. Debian's text sidesteps that by refusing the label and keeping the standard. I think that's defensible and I also think it makes the ban camp's problem worse, because their complaint was never about disclosure. It was about volume. Debian's policy does nothing about volume.

The practical move if you send patches upstream: read the contributing guide before the first line of code, not after the diff. And if your agent generated the patch, you own every line of it in the exact same way you'd own it if you'd typed it. That was true before the vote.