I have some doubts about whether we should distinguish and treat differently the use of AI code assistants as an autocomplete tool for small chunks of code (e.g. when writing a loop around features or when the same logic applied to a Red band needs to be applied to Green Blue and Alpha just incrementing the band counter by 1) vs. when AI tools are used to generate entire functions/method/classes.

The main different is probably that in the first scenario the developer would have been perfectly capable of writing the code autonomously but used AI as an enhanced autocomplete tool to speed up the development and avoid tedious (and error prone) copy-paste-replace-repeat loop.

IMHO the lightweight use of AI tools doesn't pose any additional threat in terms of IP and maintainability of the generated code because the same quality of code would have been produced by a developer who haven't used AI.

I feel the same way about LLM completions, but as they get more capable, the line between them and the usual kind of code generation blurs. Is this a lightweight use of AI or not?

Looking quickly, starting from the repo and CONTRIBUTING.md, I didn't find anything explicit about licensing, inbound=outbound, DCO, etc. In old-school Free Software, that means inbound=outbound, but I don't feel it's a safe assumption that those coming to gdal have that shared cultural heritage.

i know opinions differ about whether LLM output is infringing (I say yes), but if the project is going to have a policy other than "no LLM output allowed", I think it should first be clear on what the IP terms are for contirbutions. Maybe they are clear and I missed it.

As for the policy, same comments as for qgis: positive for demanding disclosure and prohbiting extractive contributions, and negative for allowing LLM code that I see as not coming with an adequate license grant.

coverage: 71.697% (+0.007%) from 71.69%
when pulling 3e84bcc on rouault:rfc111_ai_tool_policy
into 2acbd45 on OSGeo:master.