[gdal-dev] Integrating AI assisted code review for PR in OSGeo/GDAL?

[Kurt Schwehr]

My personal take:

I have been using Gemini based code reviews in addition to humans for more
than 6 months in other code bases. They have gradually gotten a lot better.
It started off with 5-10% of the review comments being useful. Now it's
more like 75% of the comments are useful. However, I occasionally end up
spending significant time figuring out that something that sounds right
isn't quite right for the code at hand. Or more challenging are responding
to humans who insist I address AI generated comments (I've been that person
too...)

Overall, it's been a win in getting a lot of little bugs out of code and
some of those bugs have been in the code for more than a decade. Just be
ready to ignore a lot of AI comments.

-Kurt

On Thu, Apr 2, 2026 at 6:43 AM Even Rouault via gdal-dev <
gdal-dev at lists.osgeo.org> wrote:

> Hi,
>
> I think it could be worth to have the *possibility* of requiring an AI
> assisted review for pull requests, directly available from our canonical
> repo. I've been occasionaly experimenting Gemini Code Assist and Copilot
> in my personal fork. Copilot had repeated failures a few weeks ago but
> seems to have been fixed recently, so I've more experience with /gemini
> review. I find it useful and it has spotted real issues, some of them
> would have probably went unnoticed during classic human review, and with
> an acceptable rate of false positives or debatable remarks.
>
> So my proposal would be to have the tool(s)  enabled in OSGeo/GDAL repo,
> *on demand* (not sure if that's possible for Copilot. Is that a setting?
> Although I'm not trusting github enough to be sure if we want to
> increase our use of it.  Gemini review is definitely on demand and an
> external github app we can disable in one click) for developers or
> reviewers that want to trigger them. I don't think having them to run
> systematically is a good idea, because some PRs are too trivial to get
> any benefit from them, and having them enabled systematically lead to
> noise as PR comments and notifications.
>
> I definitely don't think those tools should replace human review. AI
> tools are instructed to flatter your ego and will never say your PR is a
> bad idea, which a human reviewer will occasionally say. Or they lack the
> global picture, etc. I see them as additional tools on top of our CI
> instrumentation and human review.
>
> Anyone with experience in that area and thoughts?
>
> Even
>
> --
> http://www.spatialys.com
> My software is free, but my time generally not.
>
> _______________________________________________
> gdal-dev mailing list
> gdal-dev at lists.osgeo.org
> https://lists.osgeo.org/mailman/listinfo/gdal-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/gdal-dev/attachments/20260403/52ffb811/attachment-0001.htm>