Most of the models where people are concerned about don't do this when unquantized, so I doubt it's much about the metapolitics imposed in reinforcement training.