I assume you don't mean keyboard text predictions, which would be a different thing, but the platforms.
It's a new convenience feature. Something they as a platform can shine with, retain users, and set themselves apart from other platforms.
Having training data is not the primary potential gain. It's user investment, retention, and interaction. Users choosing the generated text is valid training data. Whether they chose similar words, or what was suggested, is still input on user choice.
It does lead to a convergence to a centralized standard speak. With a self-strengthening feedback loop.
The Web Archive "Wayback Machine" is a project from archive.org, which does much more in archiving and accessibility efforts. An alternative service for websites is https://archive.ph/.
54% of Wikipedia pages contain at least one link in their “References” section that points to a page that no longer exists.
My impression was/is that over the last years/decade Wikipedia made efforts to/switch to not linking directly but extending direct links with (dated) Web Archive links or using Web Archive links directly (dated as "sourced from this in this state; which protects against upstream edits too).