Just like I said over Starcraft and DeepMind, an AI is likely going to be worse than a person if bluffing is a big part of the game. In our system, prosecutors and cops have immunity that basically let's them lie about severity of evidence or true results of prosecutions to force deals. They have the full set of human language and imagination to use to do this.

Machines can't barely translate text between two languages or play Poker with its simpler bluffing. Anyone saying they're going to pull this off should be considered full of it until they have solid proof it's even possible. It would have to be the dumbest prosecutor on Earth and/or simplest case for one of today's AI's to outmaneuver them.