Debian developers balk at unannounced LLM reviews in NEW queue
Maintainers say automated generative-AI feedback on package uploads conflicts with a recent project vote and dumps unvetted work on volunteers.
Debian Developer Johannes Schauer Marin Rodrigues says he will no longer upload packages to the NEW queue after learning that submissions are being run, without announcement, through a large language model review workflow.
The automated reports carry a disclaimer that they were not written or checked by a DFSG team member and are not a decision on the package. Schauer Marin Rodrigues called the setup a waste of volunteer time: either maintainers read output that may be wrong, or they ignore it and still burn the compute used to generate it. He argued the practice also sits poorly with Debian's recent general resolution on generative AI, which neither endorsed nor banned the tools but said large-scale automated actions should seek prior discussion and consensus. Applying an LLM to every NEW upload, he wrote, looks like project endorsement and like the kind of broad automated change the resolution flagged.
Other developers raised similar objections. Marco d'Itri said unreviewed LLM text is not something maintainers should be asked to process, and that the DFSG team must own the output rather than disclaim it. Mathias Gibbens reported that sample reviews run from several hundred words to more than 1,500, a sharp jump in bulk compared with prior human notes, and that the process appears to penalize uploads that ignore LLM complaints while rewarding those that satisfy the model. Matthias Geiger cited concrete errors, including bad advice on bundling a library Debian already packages and a false blocker over a Build-Depends acceptable in experimental, and said the silent rollout further cut his motivation for Debian work.
Shengjing Zhu took a milder line, noting the messages are explicitly not decisions and can be filtered, and that human reviewers should still make the final call. Antonio Terceiro pointed to an open request for an opt-out. Martin Steigerwald agreed there should be no penalty for ignoring LLM notes unless a human reviewer has already judged the points legitimate.
Critics frame the change as the first major Debian-wide process shift involving generative AI since the vote, introduced without project discussion and shifting cost onto uploaders.