Re: Wikipedia at 25: A Wake-Up Call (essay)

‫בתאריך יום ה׳, 30 באפר׳ 2026 ב-18:11 מאת ‪Christophe Henner via
Wikimedia-l‬‏ <‪[email protected]‬‏>:‬
>
> Put Wikidata and Commons (and Wikifunctions) front and center as core
resources. I know not all projects are enthusiastic about them but we do
need to work that out because they are each in their own way the best way
to mutualize resources and work across projects. This needs to be a clear
goal and to find out how we make it happen properly together for each
projects.

Why is Wikifunctions in parentheses?

How can we put Wikifunctions front and center as a core resource? What does
this resource provide at the moment?

> AI for editors. We need to embrace AI, and I know this is a complex
decision.

Wikimedians usually love details and specificity. What does "embrace AI"
mean? "AI" is an elusive and practically undefinable term. We don't need to
embrace something that we cannot even define.

> AI-assisted translation across linguistic versions,

We've been doing it since 2015, we just usually don't call it "AI". I'm one
of the developers of the first version of the extension that allows this,
so I'm biased and I won't say anything about how well it works; everyone
can check and judge for themselves ;)

> AI-powered content gap detection,

Why is "AI" needed for this? And what kind of "AI"? Isn't it enough to
query Wikidata to find articles are missing and cross-check it with lists
of popular articles by country or manually-curated lists of necessary
articles? (And it *has* to be manully curated for proper unbiased cultural
relevance.)

> AI drafts for stub articles reviewed by humans,

This is already being done using ChatGPT and other
existing coversation simulators. Some language editions of Wikipedia are
trying to ban it, but it will keep happening. Bad editors will keep
publishing bad articles without proper review despite the bans, and
patrolers will increasingly (and rightly) complain about this. Or worse,
they'll burn out. Good editors will keep publishing articles whose early
version was created by an LLM after careful review despite the bans, and no
one will complain because they are carefully reviewed and the bad things
that LLMs do won't be there.

The question is what else is left to do?

Making Wikimedia's own LLM—truly freely licensed and trained specifically
on community policies? It's conceivable, possibly useful, possibly not even
very expensive to develop, though possibly expensive to host because of the
energy cost. And some people may also complain that it's not an ethical
thing to do at all for environmental reasons, and quite possibly they will
be right.

A really interesting innovation would be trying to make an LLM that is
easily fixable without having to go through a very long retraining. Another
interesting innovation is to make an LLM that can truly work with databases
rather than just being an LLM. Those innovations will be costlier, however,
and I'm not sure that they are possible at all.

> AI assistants to help newcomers to navigate rules,

The Growth extensions are not so bad at this, without much "AI". The
development is slowish because of the differences between wiki communities,
but the results are not bad.

> CI/CD for our content to make content reviewing able to absorb high
volume of AI slopes without exhausting users.

Not sure why do you call it "CI/CD", but yeah, content reviewing to prevent
slop is much-needed. I don't know how to make it well, however, and I'm not
sure that anyone does. It's a good thing to research, but my expectations
of success are low.

_______________________________________________
Wikimedia-l mailing list -- [email protected], guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines and https://meta.wikimedia.org/wiki/Wikimedia-l
Public archives at https://lists.wikimedia.org/hyperkitty/list/[email protected]/message/3KZMUWQ5ET633DY2Q53DJV7IUNFHE2QN/
To unsubscribe send an email to wikimedia-l-leave-RusutVdil2icGmH+5r0DM0B+6BGkLq7r@public.gmane.org