My gut reaction is that this won’t work long-term. Users on youtube often point to specific timestamps in a video in comments or link to specific timestamps when sharing videos, meaning there needs to be some way to identify the timestamp excluding ads. And if there’s a way to do that there’s a way to detect ads.
Of course, there’s always the chance they just scrap these features despite how useful they are and how commonly they’re used; they’ve done similar before.
I’m not sure where you’re getting the idea that language models are effective lie detectors, it’s very widely known that LLMs have no concept of truth and hallucinate constantly.
And that’s before we even get into inherent biases and moral judgements required for any form of truth detection.