Debian developer refuses NEW uploads after DFSG team silently adds LLM review
- Debian developer Johannes Schauer Marin Rodrigues says he will stop uploading anything that needs NEW queue processing, after learning that uploads are now run through an LLM based review he was never told about.
- He argues the DFSG team putting every NEW upload through an LLM conflicts with the 2026 GR, which says Debian "neither endorses nor prohibits" generative AI and calls for prior discussion and consensus before large-scale automated project actions.
- His preferred fix is an opt-in: default to no LLM review, with developers explicitly choosing to receive the generated report.
- He would rather see that time go into Lintian, which gives reproducible answers he can fix or override, unlike a model he calls a binary blob with non-deterministic output.
- In the HN thread a commenter measured two dashboard reviews at 734 words for a pass and 1,535 words for a failure, about five printed pages, and another said ignoring the LLM review pushes an upload lower in the queue.
Hacker News 의견들
Still sad that Debian is wiring an external LLM into the build path, it goes against everything the project says it stands for. There's no alternative to Debian for me, so I'm about ready to sell the machines and buy goats.
The other way is simple: don't spam people with LLM output, actually review the package yourself.
Does the NEW queue even have this problem? You can't just upload there, you have to be a Debian developer or maintainer already.
How is answering LLM slop with more LLM slop a fix? EU watermarking plus detection is the only other route I see and we are nowhere near it yet.
The lesson is that if you don't own the project, things change without you and all the hours you put in become retroactively wasted. That's a big reason I never got into open source.
I'm not anti-AI, I'm anti making people read LLM output that nobody bothered to read first. And they punish you for ignoring it by pushing your submission down the queue.
It's one group being lazy and shifting work onto another group. AI just happens to be the tool they picked this time.
I do generate LLM messages sometimes, to reply to unsolicited generated pull requests. Uploading to NEW is different, that's a pre-approved developer, not a random stranger.
What about script-generated review comments from before LLMs existed? Do you have the same complaint there?
The real gripe is communication. Use AI to help write code, then explain it in your own words. Dumping a wall of hallucinations on someone and expecting them to act on it is lazy to the point of hostile.
An opt-out would solve this. Even just a "no LLMs please" line in the upload that skips the review, while keeping default-on for everyone else.
Debian has a history of taking the hard right road even when it's slow and painful, and that's exactly why so much else is built on top of it. I'd rather they keep doing that here.
So how often is this LLM actually wrong?
Wrong question, and developers shouldn't have to burn their own time finding out. Nobody even knows what the setup is, local 8b qwen or paid claude behind it, none of it is documented.
Picked two reviews off the dashboard at random: the passing one was 734 words, the failing one 1,535 words, roughly five printed pages. Expecting someone to swallow five pages of slop to learn why their review failed is rude.
No Linux distro can hold a hard anti-AI line. All the BSDs do, mostly because you can't prove what license the model's output carries, so it can't legally go into BSD-licensed code.
Some developers just don't get that shoving AI in people's faces is rude. iTerm2 now pops sloppy AI images in menubar tooltips. What happened there?
Useless hero images for tooltips, and being slop is just the cherry on top.