The conclusion I came to is that the effort spent writing the various stretchtext possibilities, and it is quite considerable effort, is better spent just making your writing better in the first place.
I think that will be true regardless of the human-based implementation.
AI shrinking tech might work, but even so, it's a delicate operation. It's so easy to drop a "not", or a more complicated equivalent, that makes the summary not actually what you said, and if we're back to "human reviews the AI attempt at shrinking your text" we're back to "the effort is probably better spent improving your writing".
There's also a problem where assuming a decent reader, the cognitive load of deciding whether to stretch the text, performing the requisite UI action, re-acquiring the text you just manipulated and re-reading it, is not entirely dissimilar from the cognitive effort of just reading the full text in the first place. By the time the amount of text you're hiding is worth all that, you probably should have linked to a real page, or just used a footnote, which is basically a centuries-old tried & true mechanism for "stretch text" that everyone knows what to do with.
Relative to the benefit, it's a lot of cost on both producer and consumer.