[gdal-dev] Proposed RFC 8 amendment regarding (prohibited use of) generative AI tools

[ElPaso]

Il 09/10/24 00:55, Greg Troxel via gdal-dev ha scritto:
> ElPaso via gdal-dev <gdal-dev at lists.osgeo.org> writes:
>
>> I have read the discussion on lwn and I must say that I am more in
>> line with the debian position.
> My view is that code that comes out of generative AI should be viewed as
> an improper derived work, and lacking adequate provenance/permission to
> be added to an open source project, period.  As Even says, we could
> relax this in the future.  It's very difficult to go back and remove
> things.


This may be true in some circumstances but that's not what I have seen 
so far using copilot wih GDAL: most of the times what it does for me 
much faster than me is cut-and-paste-replace or autocomplete taking code 
from other parts of GDAL or most frequently from other parts of the same 
file that I am editing, for example, when I was changing  ogrlineref to 
use gdalargumentparser, after manually changing the first part of the file

https://github.com/OSGeo/gdal/pull/10147/files#diff-b906434b9e6a52aef54e0894ba43a7202290b1af4964cc5d9f1ec8ae7a1c4e15R1271

the AI was very useful to autocomplete the other command line switches 
(one by one, not all of them), I had to change/edit almost everything 
but the scaffolding was there, the source of the autocomplete was 
obviously GDAL itself.


Where is the copyright issue in this use case?


Another situation where I find it useful is when writing tests: most of 
the times tests are boring boilerplate code to construct the test data, 
for example: the next line here was generate from the comment

https://github.com/qgis/QGIS/pull/58986/files#diff-8eebd3707bdc54c42ecd1a2abfca8f623a4c6a8c44c89f344a6762fe95a95059R4682 


also the checks were automatically generated after I entered the first 
couple (they are essentially a copy-paste), the source is QGIS itself.


That said I agree that when an AI will be smart enough to be able to be 
"creative" (I know, hard to define what it exactly is, but a one-line 
cut-paste-replace from the same code base certainly isn't) we will have 
a problem to define who/what is the author of the code.


Perhaps we could find a way to allow the limited use of AI tools as 
autocompleters as long as the source of the generated material is 
obviously the code base itself (for instance when using the GDAL API).


-- 
Alessandro Pasotti
w3: www.itopen.it