There is an interesting concept of a very smart AI but not capable enough of taking over and knowing this making a deal with humans to prevent an even smarter AI taking over, by for example monitoring training runs and reporting on rogue AIs.