i think those things just like the ai things are just probability based. which word, token or letter is most likely, based on occurences seen in large data sets.
thats not to say prediction is not a thing, like forecasting. but thats also generally based on large datasets / historical data, and perhaps as much 'context' as you can get.
context being whats around the thing u want to predict thats related to it in some way.
That's character-level n-gram modeling
I always assumed that was just using spell-check in a unique way. I could absolutely be wrong about that, though.
What I'd probably do is to start scouring the literature for "predictive text algorithms" and see where that rabbit hole leads.