If the bottom line are donations - as the article states - why push for getting AI companies to link people to Wikipedia instead of pushing for the companies to donate?
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.
If they destroy the relatively high-trust internet, the low-trust replacement will require digital ID for every client, with non-neutral traffic price varying by {business digital ID, content}. No more free geese, even to check whether there is a golden goose worthy of payment.
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.
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.
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.
