A less drastic extension is Clickbait Remover. It replaces thumbnails to an actual frame from the video and de-capitalizes titles: https://addons.mozilla.org/en-US/firefox/addon/clickbait-rem...
My thinking is that with curated thumbs I can at least get a good idea of the author's intent (particularly when one's a relatively experienced regular YT user).
I kinda thing targeting curated thumbnail selection is looking at the wrong cause: curated thumbnails don't cause clickbaity thumbnail choice, perverse commercial incentives do.
I'd like to think I could select videos on title alone but if that's your route then just remove rather than replace thumbs (and text can still mislead).
To put it another way: even if we agree the current thumbs are bad (annoying, misleading), are the replacements better (or just irrelevant). Most people like to judge things on intent, but intentionally bad is still better than unintentionally worse.
And in the end, I can much more easily develop intuition to spot intentionally deceptive common human patterns than spotting unintentionally irrelevant random automated patterns.
Older YouTubers used to manipulate this by putting the suggestive part of the video at the right time.
Here's a blast from the past doing just that: https://www.youtube.com/watch?v=r8tXjJL3xcM
And if the thumbnail is picked randomly, you are just rewarding videos that show dramatic content with text overlay at all times. And the best-researched Tom Scott video might get randomly punished because the thumbnail is a brown-grey blur in the middle of a transition.
The existence of the medium "thumbnail" is not more of a problem than the existence of book cover. Breeding a healthy culture of what should be on it is the problem we should focus on, both by audience behavior and suggestion algorithm.