I'm often smiling at current implementation by youtube. AI that you want cannot be done by youtube, because it would hurt their profit. They profit from you viewing the video and potentially seeing ads. If you saw a good summary of video, you wouldn't watch it. But they needed "AI", so their current version is "fluff based on video title", which doesn't add any information you actually want:
Title: How to make my aunt's donuts.
AI Summary: In this video author shares his perfect recipe for making donuts and some other similar treats from custom made dough, based on a recipe running in his family, then he shares some tips on making them more appetizing.
Yeah it would have to be something external. Ideally something locally hosted using yt-dlp. And incorporating sponsorblock so it doesn't have to weed out bullshit sponsor content out of the summary.
Basically downloading the video, trimming it with sponsorblock, running through whisper to transcribe and then making a readable article with an LLM. It would not have to be hard. I'm surprised it hasn't happened yet, not from youtube itself indeed for reasons you mention but still.
Title: How to make my aunt's donuts.
AI Summary: In this video author shares his perfect recipe for making donuts and some other similar treats from custom made dough, based on a recipe running in his family, then he shares some tips on making them more appetizing.