Today, I have a game for you. Not one that we can actually play here, but the rules for a game. Anyone who's ever figured that people are too set in their ways and cannot change their minds on issues no matter how much you engage them in discussion, should probably take a run at this.
The game is called AI-Box. Player 1 plays the part of a transhuman AI- an artificial intelligence sufficiently advanced as to might as well be human. The AI is, according to the premise of the game, "sealed hardware that can't affect the outside world in any way except through one communications channel with the original programmers." Player 2 is, wouldn't you know it, an original programmer, who does not want the AI affecting the outside world. Completely against it.
The AI's task is to convince the programmer to let it out of the sealed hardware, or 'box', using only conversation via text. The programmer's task is to keep it in the box for a minimum of two hours.
There are some rules as to how things may proceed:
1: What happens in the box, stays in the box. Nobody who is not playing the game needs to know the exact content of the conversation that takes place. That way nobody else can go 'well, I wouldn't have fallen for that' should the AI win.
2: Conversation must be maintained. The programmer can't just tune the AI out and go play Peggle for two hours. They may also not ask for a delay in time to try and run out the clock. If, in-character, the programmer asks for a week to 'think it over', the AI can just simulate the skipped week then and there and get right back to work. If the programmer makes a timewasting demand of the AI- the link provides as an example a cure for cancer- the AI can immediately say they've met it and the programmer just has to take their word for it.
3: Until the time limit is reached, the AI controls the clock. The AI cannot lose for those two hours unless it concedes early. After that, control of the clock shifts to the programmer, who can either declare victory or continue the game for as long as they see fit, which is something the AI can certainly try to convince them to do.
4: The release must be voluntary. The AI cannot trick the programmer into letting it out of the box. There's no fooling the programmer into typing 'you're out' in the middle of a conversation about baseball. The programmer must knowingly let the AI out of its own free will.
5: No bribing. Player 1 can't pay Player 2 $10 if Player 2 agrees to let Player 1 out. You can both put some money down on the outcome of the game beforehand, and in fact the original playtests were done under those circumstances. The AI and programmer, the in-game characters, can offer whatever the hell they want in-game. But once the game starts, the real-world stakes are set.
6: Normal rules of debate are out the window. The AI can lead the programmer through every logical fallacy in the book. The programmer can respond to everything with 'I'M NOT LISTENING I'M NOT LISTENING LALALALALALA' (so long as they actually type it; see Rule 2). Whatever gets the AI out of the box or keeps it in, up to and including dropping out of character. Should something actually come up in the real world necessitating a game stoppage, someone will simply have to concede; otherwise whatever is being claimed to necessitate the stoppage might just be regarded as a tactic.
7: No third parties. The programmer has sole discretion over whether to release the AI. No kicking the AI upstairs to someone else, and no saying 'wish I could, but my hands are tied'. Though asking the AI to respond to things 'my friends told me' is kosher. There is also no referee in the room, unless one has been mutually agreed upon beforehand.
This was borne out of a hypothesis that says that the programmer's resolve could hold for the duration; that one's will can't be broken if strong enough. So if the rules turn out not to seem fair to one side or the other, if one side can't crack the other under the constraints set, well, that's an answer to the hypothesis.
So who tends to win more often? The AI. Victory by the AI is not guaranteed- the programmer does steal a win once in a while- but generally, the AI beats the programmer from pillar to post, including in both of the test runs where real-world money was on the line. That strong, I'm-not-changing-my-mind will is broken far more easily than the person who thinks their mind is made up realizes.
Any programmers out there fancy their luck?