I have a theory about current state of the are of recommendations, automation or ML in general. I think a lot of the valid criticisms we see with recommendation

I have a theory about current state of the are of recommendations, automation or ML in general. I think a lot of the valid criticisms we see with recommendation systems (building of bubbles, extremization etc.) is due to the simplicity of the algorithms. In reinforcement learning there is a notion of exploration and exploitation. All of supervised ML can in a way be seen as exploitation. Incorporating ideas of exploration in a clever way into the standard recommendation algos, I think could be the solution to a lot of the mentioned problems.