Debian is holding a General Resolution on whether LLM-assisted contributions belong in the project at all, and the most striking part of the proposal text is that the hardest objection isn’t quality — it’s provenance. Proposal A wants an outright ban on any contribution written with generative AI, and its first argument is copyright: LLM output has no clean license lineage, and Debian’s policy requires absolute clarity on licensing before anything ships. Human-authored code with murky copyright is already rejected. The proposers ask why machine-authored code should get an exception.

That framing is the one I keep running into on the production side. Everyone debates whether the model’s code is correct; almost nobody has a clean answer for whether it’s ownable. In an enterprise pipeline you can bolt on review, tests, and eval harnesses to catch the accuracy problems the proposal lists — watch files that don’t work, overrides pasted out of context, invented copyright headers. What you can’t easily bolt on is a provenance record that says this artifact is safe to redistribute under your license. Debian is a distributor by definition, so the licensing question is existential for them in a way it isn’t for a team shipping an internal service.

The community argument is the sleeper. The proposal claims LLM-dependent new contributors never learn the packaging craft, so they can’t grow into the maintainers who replace burned-out ones — the review burden goes up while the bench gets thinner. That’s a governance problem dressed as a tooling problem, and it maps directly onto how agentic coding changes who becomes a senior engineer. The HN thread splits hard on it.

If a project this careful can’t write a workable disclosure rule instead of a ban, what does that say about the “just require attestation” plans the rest of us are counting on?