Problem with AIs is to define "good" in this case. When do you call it a day? What is a "good" thumbnail?
Of course you could give an AI thousands of thumbnails in relation to the video title (clean your data and be careful of the quality) and train it on said dataset. Afterwards you could give the AI the title and possible an image input, after which it might edit it / include texts; but I think it's highly unlikely that this will look
ANY good unless you've got a professional team covering this project.
If it works, I'll happily see it at "Two Minute Papers" the other day