Re: GR -- Allow AI-Assisted Contributions

Hi,

On 2/24/26 23:14, Theodore Tso wrote:

> Some anti-AI voices are concerned that use of AI will decrease the
> ability to gian seasoned contributors, with the implied concern that
> this is self-defeating because it restricts the ability to gain new
> members in the future.  And you are now saying we should gate keep
> contributors that might be using AI as being unworthy of contributing
> to Debian?  I'd say that is even more self-defeating.

I'm saying we need to define criteria:

  - Should AI-generated answers and essays be accepted in the 
"Philosophy and Procedures" part of the New Maintainer process?
  - Are we accepting candidates who are entirely reliant on AI to 
perform simple tasks?
  - If no, how do we distinguish between "might be using AI" and "is 
entirely reliant on it"?

There absolutely is a level of "using AI" where people do not have the 
necessary skills to adequately review the AI output, and where the best 
possible outcome is that they are burdening other people with these 
review tasks, and the worst outcome is deploying changes without proper 
review, causing other people to have to step in and do damage control.

As I said, from a technical perspective[1] I have little concern about 
seasoned contributors using a tool and reviewing its output, but free 
software projects can only accept these contributions if the cost of 
doing so is low enough, and that largely boils down to an established 
trust relationship and long time tenure in a project.

For newcomers, we still need to find someone experienced to review their 
large changesets, and that, too, is a cost for the project.

> We already probably have developers in Debian, that are using
> closed-source IDE's because when Java programming they don't want to
> have to type long variable or class names, such as
> InternalFrameTitlePaneMaximizeButtonWindowNotFocusedState.  Shall we
> run them out of Debian because they are using non-free tools, even
> though the output might still be free?

I have zero problem with deterministic tools. Replacing the data type of 
a variable and then following all expressions where it is ever used (the 
example you give in your other mail) is a common enough problem that 
such tools have existed for a long time. I remember using Visual Assist 
for that twenty years ago, and my expectation is that these tools still 
hold up a lot better than any AI solution today, work offline, and cost 
less for a perpetual license than the token cost for one invocation.

On the topic of token cost, I fully expect that to "solve" the problem 
of AI contributions in the near future anyway.

    Simon

[1] usual reminder that the externalized cost of this technology is 
still too high.

-----BEGIN PGP SIGNATURE-----

iQEzBAEBCgAdFiEEtjuqOJSXmNjSiX3Tfr04e7CZCBEFAmmkEG4ACgkQfr04e7CZ
CBEwAggAlKuZMkkzk7S0gNZgMCZUzV2Vf/Zx7g5V7gPDPsrFreiJWEYL+3r+TOLH
xgRcjKHk9egIgAF73uNRyXS5HBDj/CEcAsiWmDT/iSUtSWHFlTXfT3JKAP/X8CVe
7xaGr4FtHX8PcqhekM2BMAECAEg6E4LHeYHHMublNPBb1PNC7ngnViAumeSpTlMk
bE6EJlmUyKo78If1T/HLiDVYrqFRr4WdQkyDqn9QlMFtv93vo3USsJPgX7W19CkS
U/b+Bec3iCN8bI6dhlPlDPeWzjirGWxk73MkNt22/SemK74udkwZTWAqYxbjx2yM
yNzL+vIAaH4sSLYtRMXdqlM0IQqzEg==
=EJLg
-----END PGP SIGNATURE-----