The easiest way this could work would be allowing you to group your subscriptions manually or automatically tagging them into preset categories (which would still use AI technically). It’s not really feasible to create an endless number of shitty algorithms that probably won’t give you exactly what you want.
PSA: on youtube, you can turn off the algorithm by turning off watch history, so you will only see your subscriptions page, but also related videos under the current video you’re on. I spend less time there now, and enjoy the content I do watch much more.
I only skimmed the article, but I’m familiar with this area of work. It’s the fine tuning phase of llms that require human “graders” in a sense to tell the model what is a helpful vs unhelpful response and elevate it from a literal autocomplete tool to a chatbot. The biggest platform for this kind of work is DataAnnotation if you want to look it up. It’s not just random people in the global south but also professionals training the model on their niche expertise (e.g., lawyers, accountants, engineers). The other type of work that has been going on even pre-llm chatbot era is, as you said, moderation. You moderate the chatbot for “bad” answers, correct them, and often that means impersonating the bot to the end user. I believe this is the main driver behind llm progress lately; literally humans teaching AI how to appear more human.
In this scenario, everyone is so focused on the gun that they forgot murder is already illegal. We do need gun and AI regulation, but we can’t even put the AI aspect aside for one second to investigate theft and defamation which we do have laws against regardless of the tools used.