Le Mon, Nov 09, 2026 at 06:10:44PM +0100, Marc Haber a écrit : > On Mon, Nov 09, 2026 at 05:27:48PM +0100, Matthias Geiger wrote: > > The main point isn't whether this has consequences for NEW, but rather > > that I get a non-human message I didn't ask for. And this is very > > different from mass bug filings: Those need to announced on -devel, > > and then after a discussion those are sent. However, those provide > > value, such as coordinating a transition or removing outdated > > dependencies from the archive. Those reports serve a goal to simplify > > and coordinate work across the archive. An LLM "review" of my copyright > > does not. It it just a wall of text from a model, that provides no value > > for the uploader. I can just run licensecheck locally and can get the > > same result. > > Just for the record, I'd consider the feedback I get without having to do > anything of value. Not as valueable as a feedback from an expert human > reviewer, but we don't have that kind of personpower.
The problem is that any LLM feedback is adversarial by nature. LLM are designed to elude scrutinity, to pass from a competent human. Here lie the real danger. Lucas [1] posted an example of a LLM review. The review was completly wrong but it was sufficiently plausible that Lucas and most other debian-devel commentors accepted it as valid. Some of them, including Lucas, have years of experience as DD. This failure to spot the blatant errors show that nobody is immune to LLM influence. Such behaviour by an expert human would be considered as a deception attempt and lead to sanction. An unexperienced human would not be nearly as convincing as a LLM. Do you accept to be held responsible from the action you perform following a LLM incorrect advice that you did not ask for explicitly ? The second risk is, as developers get used to follow incorrect LLM feedback without questionning it, they will start to implementing it without waiting for the LLM, and then Debian policies will be dictate by LLM hallucinations. We need a way to disable quickly any LLM that produces incorrect advices before it spreads. [1] https://lists.debian.org/debian-devel/2026/10/msg00290.html https://salsa.debian.org/dfsg-team/dfsg-omp-review/-/work_items/26 Cheers, -- Bill. <[email protected]> Imagine a large red swirl here.
