Repository navigation
Suggested elaboration of AI policies #1777
Description
Activity
A different proposal was opened: #1778
Reacted by Tim PetersSo I also propose that the docs encourage disclosure of material AI-generated content in contributions. Not at all to stigmatize,, but to help reviewers focus on the special risks that AI aids bring.
Does disclosure really protect anybody from anything?
A counter proposal is to provide LLM PR and issue templates as well as prompts and guidelines for LLMs to make sure they can do their job perfectly, and submit code that 100% complies with what community wants.
I am not sure that the perfect guideline/prompt should be. "Do not use code from another projects" or "please, check that generated code is not 100% derived from the projects with incompatible distribution rules", but that's a good start.
I would also include in the template fields to reference specific LLM that is being used, inference engine and agent prompts. Especially if it is Open Source community maintained models. Ideally, the PR should have all context to feed to another LLM to regenerate the change from scratch (and compare/benchmark the result to choose the best one).
Thanks for the feedback! I don't claim this would "fix" anything. Encouraging a culture of disclosure is about establishing social norms for those of good will to voluntarily adopt, not a hammer of any kind.
I don't think more would help. For example, LLMs are fast-moving targets, and even a single model can give materially different responses on different days. Why? Who knows? Perhaps a randomized component, perhaps an update to the training database, perhaps because a paid version sees that you're approaching your limit for "think time", ..., on & on. Reproducibility just won't be achieved.
In any case, this issue is getting so little traction that I'm going to just close this. @StanFromIreland earlier pointed to a related issue that is getting traction: #1778
Reacted by Anatoli Babenia
Describe the enhancement or feature you would like
I would like to add text emphasizing that use of AI tools substantially intensifies the chance of submissions unintentionally including derivative work (violating others' copyrights and/or licensing terms). While experienced devs should already be aware of that, contributors come from many backgrounds, and the docs should spell things out more clearly. Before bot days, one person (who later went on to become a core dev) started with an elaborate patch that included a comment plainly stating that what followed was copied from glibc source. Not malicious, they simply didn't know any better. They weren't yet a "software geek", but quite capable in the subject area.
So I also propose that the docs encourage disclosure of material AI-generated content in contributions. Not at all to stigmatize,, but to help reviewers focus on the special risks that AI aids bring.
Suggested text is in this dpo post:
https://discuss.python.org/t/i-am-concerned-about-llm-code-in-python/106691/77
Describe alternatives you have considered
No response
Additional context
No response