My guess is that the scraping tools are specialized for web, and creating per-application interfaces isn't cost effective (although you could argue that scraping Wikipedia effectively is definitely worth the effort, but given its all text context with a robust taxonomy/hierarchy, it might be non-issue.)

My other thought is that you don't want a link showing you scraped anything... and faking browser traffic might draw less attention.

The rationale I've seen elsewhere is that it saves money. It means you don't need to go to the effort of downloading, storing and updating your copy of the database. You can offload all of the externalities onto whatever site you're scraping.

Sheer laziness?

Money. One requires you to use your hardware and your developers. The other way doesn’t.

i tried doing that in summer 2019, and the downloaded formats were at that time proprietary and depended on decoders which were like a tail recursive rabbit hole.

in contrast, letting their servers render the content with their proprietary tools yields the sought data, so scraping might be a pragmatic choice still.

Because that would probably require extra work, why do that if it already crawls and scrapes it in the first place

It's possible that they don't know. I literally didn't know there was a full downloadable db until right now.

Because many small donations from individuals are better than few big ones from corporations for the independence of Wikipedia? Eggs vs baskets etc...

Wikipedia is not a company.

They should mainly be worried about their reliability and trustworthiness. They should not worry about article length, as long as it's from exhaustiveness and important content is still accessible.

Serving perfectly digestible bits of information optimized for being easy to read must not be the primary goal of an encyclopedia.

By the way, "AI summaries" routinely contain misrepresentations, misleading sentences or just plain wrong information.

Wikipedia is (rightly) worried about AI slop.

The reason is that LLMs cannot "create" reliable information about the factual world, and they can also only evaluate information based on what "sounds plausible" (or matches the training priorities).

You can get an AI summary with one of the 100 buttons for this that are built into every consumer-facing product, including common OS GUIs and Web browsers.

Or "ask ChatGPT" for one.