Here's the full talk[0] from a Microsoft lead researcher that worked with the early / uncensored version of GPT-4.
Simplified, tuning it for censorship heavily limits the dimensionality the model can move in to find an answer which means worse results in general for some reason.