[OSGeo-Discuss] AI slope reaching our projects

[Miles Fidelman]

Yet another example of the agile mentality - that one doesn't have to actually understand the problem at hand, or practice any serious engineering discipline when it comes to software.

Moving fast & breaking things, with no aforethought, is bad enough when practiced by human code monkeys. When we turn automated code monkeys to throw shit against the walls, to see what sticks - with ill-defined directions, and no review - that's the end of civilization.

We need to start applying serious subject domain expertise, and serious engineering practice to the development of software - doing the math, serious systems architecture & engineering practice, serious design reviews from the specification stage onwards.

Otherwise it's game over for our human family enterprise.

(Then again, if we're going to start releasing autonomous Skynet drones into the world - maybe we WANT to program them with code that doesn't work at all.)

Miles Fidelman



________________________________
From: Discuss <discuss-bounces at lists.osgeo.org> on behalf of Even Rouault via Discuss <discuss at lists.osgeo.org>
Sent: Saturday, April 11, 2026 9:14 AM
To: OSGeo Discuss list <discuss at lists.osgeo.org>
Subject: Re: [OSGeo-Discuss] AI slope reaching our projects

Hi,

So 2 months after having written be below message, and GDAL, PROJ, QGIS
having adopted this human-in-the-loop policy, I personally consider it
as mostly a failure (which was a very likely outcome). I've no idea
about how much seriousness contributors using AI tools give in reviewing
their output, but the reality is that even if they are well intended and
super careful and read every line and (believe they) understand what it
does, they just lack the experience to have a critical eye (which is the
reason in the first place why they needed to use it). The asymmetry of
effort between the ease for them to push the PR and the ones bearing on
the reviewer's shoulder to debunk the flaws is unsustainable. Pretty
similar to https://en.wikipedia.org/wiki/Brandolini%27s_law . With the
difference that it is far from being obvious where the "lie" is (which
is expected since those tools are designed at generating credible
output). You've have to be super mindful of the tendency of LLMs to
generate bloat. Unit tests are a striking examples. Before AI, we were
begging our contributors with "please write a couple unit tests,
please". No this is more "please please cut down the tests to the bare
minimum, and remove all those comments that bring nothing but increasing
cognitive burden". Crazy!

So my mindset would be now more on the side "no AI generated code
allowed", but I don't know how we can realistically enforce that. I bet
a significant proportion of people addicted to those tools will for sure
lie (addiction is the appropriate word), which will make things even
worse if you don't know when you have to turn your
review-with-extreme-scepticism radar on.

Any feedback from other projects that the ones mentioned here?

Even


Le 05/02/2026 à 18:14, Even Rouault a écrit :
> Hi,
>
> At least GDAL, GRASS-GIS and QGIS have seen recently a raise in AI
> driven "contributions". That's really a plague. GRASS-GIS has adopted
> a policy regarding that in
> https://github.com/OSGeo/grass/blob/main/CONTRIBUTING.md#ai-use-policy
> . QGIS is about to have one in
> https://github.com/qgis/QGIS-Enhancement-Proposals/pull/363. GDAL will
> likely do something similar soon.
>
> Like there's a Code of conduct that projects can use, maybe OSGeo
> could have a similar thing. Although I suspect people may have
> different sensitivities to the subject.
>
> Even
>
--
http://www.spatialys.com
My software is free, but my time generally not.

_______________________________________________
Discuss mailing list
Discuss at lists.osgeo.org
https://lists.osgeo.org/mailman/listinfo/discuss
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/discuss/attachments/20260411/2235f770/attachment.htm>