Could we create AI societies with checks and balances? AI workers do tasks, AI cops (smaller, simpler models) monitor for misbehavior, AI prosecutors assemble evidence, AI defense agents challenge accusations and look for false positives, AI watchdogs inspect the cops and judges and we could have competing institutions using different models/training methods/owners and incentives.

Specifically, I think the AI cops, prosecutors and defense agents can be highly specialized and less intelligent models than the AI workers as they only have to understand this more limited domain of having the workers stay within the bounds of the task. Because these are more specialized and less intelligent, we humans could feel more confident that these are aligned.