I believe that the number one way to combat the radicalizing effect of recommender engines would be to establish new, more human KPIs rather than the current KPIs used today which are likely either time spent on the platform or number of clicks. However, neither of these metrics reflect the interest of the user - they both reflect the interest of the platform and maximizing advertising revenue. This to me is what is most lacking in large commercial recommenders: a human element that advocates for the user first.
The recommender does not need to only use the user-focused KPI, it can balance the needs of the user with the needs of the platform by using some weighted combination of metrics. Designing the user-focused metric is not a simple task and there are many possible approaches. Perhaps the simplest would be to ask the user: are you happy with the recommendations? However, this metric may suffer from users not answering or from users how are subconsciously affected by recommendations and are still satisfied.