Basic AI Contribution Policy

[sirosen]

I'm planning on putting in place a basic policy on LLM usage for check-jsonschema. There have only been a few slop PRs so far, but they are persistent, annoying, and sometimes demoralizing -- I doubt they'll go away as an issue.

cc @Julian, if you like, I can do a first iteration there, and if you like it you can refine it into a policy for other repos or the whole org. I'm happy to discuss this as much or as little as you like.

There are various policies in use today, at various levels of strictness. I'm willing to discuss in greater depth, but the basic principles I plan to set down are:

• I can't police what other people do, so I won't.
• Code and prose should be short and precise, not verbose and convoluted.
• Ownership and responsibility are important, regardless of what kinds of tools are used.
• Fully autonomous code submissions are not productive and are banned. (Detecting these is a moving target, but sometimes we know.)
• Don't let a machine think for you, and don't let it talk for you. It's okay if your writing is not "polished".
  • Translation is understood as a special case. Even so, if contributors are comfortable writing in English, at whatever level they are able, that is preferable.

[Julian]

It's a good topic for sure.

There have only been a few slop PRs so far, but they are persistent, annoying, and sometimes demoralizing -- I doubt they'll go away as an issue.

For the main repo it's honestly bad enough that I barely look at PRs from anyone who isn't a prior contributor at this point,so I essentially don't accept PRs from new contributors unless it's trivial to tell they're a "well-regarded human" whatever that means.

I can do a first iteration there, and if you like it you can refine it into a policy for other repos or the whole org.

Certainly OK with this. The policies I keep handy in case helpful are:

• Linux Kernel
• CERN
• LLVM
• Recurse Center
• Monarch
• Jellyfin
• Contributor Poker and Zig's AI Ban
• Have a Coherent AI Policy on [[leadership]] and touching on [[software metrics]]

but certainly this is an... evolving topic and I'm sure there are a ton we could draw from.

I agree strongly with all your bullets as well so even just those seem like a win!