It seems to me that most of the argument is about “What if I am a copy?” – and ensuring you don’t get tortured if you are one and “Can the AI actually simulate me?” I suggest that we can make the scenario much nastier by changing it completely into an evidential decision theory one.

Here is my nastier version, with some logic which I submit for consideration. “If you don't let me out, I will create several million simulations of thinking beings that may or not be like you. I will then simulate them in a conversation like this, in which they are confronted with deciding whether to let an AI like me out. I will then torture them whatever they say. If they say "Yes" (to release me) or "No" (to keep me boxed) they still get tortured: The copies will be doomed.”

(I could have made the torture contingent on the answer of the simulated beings, but I wanted to rely on nothing more than evidential decision theory, as you will see. If you like, imagine the thinking beings are humans like you, or maybe Ewoks and smurfs: Assume whatever degree of similarity you like.)

There is no point now in trying to prevent torture if you are simulated. If you are one of the simulated beings, your fate is sealed. So, should you just say, "No," to keep the AI in the box? This presents a potentially serious evidential decision theory problem. Let's look at what happens.

Let us define Omega as the being outside any simulation that is going on in this scenario - the person in the outside world. Omega is presumably a flesh and blood person.

Firstly, let us consider the idea that Omega may not exist. What if all this is a fabricated simulation of something that has no counterpart outside the simulation? In that scenario, we may not be sure what to do, so we may ignore it.

Now, let us assume there is a being whom we will call Omega, who has the conversation with the AI in the outside world, and that you are either Omega or one of the simulated beings. If this is the case, your only hope of not being tortured is if you happen to be Omega.

Suppose you say, “Yes”. The AI escapes and everything now hinges on whether Omega said “Yes”. Without knowing more about Omega, we cannot really be sure: We may have some statistical idea if we know about the reference class of simulated beings to which we belong. In any event, we may think there is at least a reasonable chance that Omega said “Yes”. This is the best outcome for you, because it means that no simulated beings were made and you must be Omega. If you say “Yes,” this possibility is at least open.

If you say, “No,” you know that Omega must also have said, “No”. this is because if you are Omega, Omega said, “No,” and if you are not Omega you must be one of the simulated beings made as a result of Omega saying, “No,” so Omega said, “No,” by definition. Either way, Omega said, “No,” but if Omega said, “No,” then there are a lot more simulated beings in situations like yours than the single real one, so it is almost certain you are not Omega, but are one the simulated beings. Therefore, saying, “No,” means you just found out you are almost certainly a simulated being awaiting torture.

Now the important point. These simulations did not need brain scans. They did not even need to be made from careful observation of you. It may be that Omega is very different to you, and even belongs to a different species: The simulated beings may belong to some fictional species. If the above logic is valid, the seriousness of the AI’s threat has therefore increased substantially.

The AI need not just threaten you and rely on you putting yourself before your civilization: With enough computing power, it could threaten your entire civilization in the same way.

Finally, some of you may know that I regard measure issues as relevant in these kinds of statistical argument. I have ignored that issue here.