The dictator can deal with this problem "can the AI be trusted" in a similar way to how he's dealing with the "can my underlings be trusted" problem. Dictators

The dictator can deal with this problem "can the AI be trusted" in a similar way to how he's dealing with the "can my underlings be trusted" problem. Dictators let people spy on their direct underlings and report any suspicious activity to them. What dictators need is good loyalty tests. Something like that is impossible for humans, you can't read thoughts. And even if someone is perfectly loyal at time A, they can always change their mind. Source code on the other hand can be read. Surely, for this the people who check the AI need to be competent and loyal to the dictator. But their loyalty only needs to exist for the duration of the check. Compare that to the general problem of having to trust people in your government for as long as your government exists.