Yes, there are signals in human behavior that you can feed to your model. But no, it is never going to learn "on Mondays he works on his comics, so he'd prefer to see webcomic-related content" Don't sprinkle in your explore/exploit experiments. I know what I want, just let me decide based on what I'm in the mood for!
TikTok-style feeds are the absolute worst offender here, where they couldn't care less about what you think. They will serve you content, and you will either say "yes" or "no". So the only option for you as the user is to just wander through their content hyperspace. There's no structured way to jump between topics because everything lives in this formless content soup.
The other problem is that many social media platforms have you subscribe to the content streams of individuals directly. Individuals are high variance. How can you teach these engines that "I only care about this person's posts about pianos, not their terrarium hobby."
I pay a lot of attention to classes of things I don't want to see at all.
If they think that the length of time I spend on a post is 100% correlated with the amount that I want to see it, they really don't understand people at all.
It's a trade off. Viewing rates give back feedback 100% of the time. Asking users for a thumbs up or down gives feedback almost none of the time, and still might not be accurate.
For example, I might see a some specific word or pattern of words in a tweet and quickly skip to the next one. That's very low friction but also a powerful signal. There are drawbacks (e.g. the long boring video that looks like something is about to happen but never does), but that's mitigated by combining this with other signals.
With 2, it's an explore/exploit trade-off. You try to explore as far and wide as you can while trying to avoid the things the user may dislike, all while sprinkling in just enough of the stuff that you know they'll like.