On 2026-11-10 06:28, Jonas Smedegaard wrote:
Quoting Richard Laager (2026-11-09 19:16:57)What if the emails were simply the output of lintian and other deterministic copyright analysis tools, would that be a problem?Yes, it would still be a problem if the DFSG team decided to spam all uploaders to NEW queue with content noone could be held accountable for.
What exactly does "noone could be held accountable for" mean?I think the DFSG team is accountable for the automated review output they are sending, in aggregate, regardless of whether the tool is deterministic or probabilistic.
If the DFSG team added `lintian`, `dnq author-check`, etc., and the tools were "mostly correct", that seems helpful. Uploaders could address the feedback, which would speed up the review of their upload. Note that address might be saying, 'Tool X is wrong here because Y.', e.g. the whole thing (that I disagreed with, but a different developer could feel differently about) 'That SUSE copyright is for an RPM file that isn't used as part of the build.'.
If the tools are getting things wrong to the point that they are creating more false positive work than helping, then the DFSG team would be responsible for that. I think it's fine for them to run such an experiment in good faith, and then adjust based on real-world results. If they fail to adjust, then they can be held accountable first socially, then through more formal mechanisms (e.g. the DPL can intervene, up to and including changing the delegation).
I don't see that LLMs change this. If they added an LLM tool and it's helpful, great. If they added an LLM tool that's producing too much bogus output, then they need to adjust.
From what I can see, the output from this specific setup is actually helpful. Granted, I've only looked at a couple of reviews that were referenced here, but those were the examples given by people criticizing the tool.
No, it would not be the same kind of problem, because there is less tension in using deterministic tools rather than probabilistic ones, partly because there is some degree of accountability in such tools (assuming they are not only deterministic but also DFSG-free) and partly because - as we have seen in the preparation of a ballot list for the recent GR vote - there are multiple dimensions involved in some probabilistic tools (including the one used here) that are difficult to keep apart¹. - Jonas ¹ I am guilty of messing up at least one distinction between reducing energy consumption and favoring green energy.
Energy consumption or other concerns about LLMs specifically are reasonable things to be concerned about. But we voted and we collectively decided LLMs are okay for this project. People can dislike the result, but it's not fair for a bunch of people to pile on the DFSG team for using a tool we explicitly decided it is okay to use.
-- Richard
OpenPGP_signature.asc
Description: OpenPGP digital signature
