>Given that the state-of-the-art literature on AI containment answers in the affirmative (AI is uncontainable in the long-term) That's orthogonal to the simula

>Given that the state-of-the-art literature on AI containment answers in the affirmative (AI is uncontainable in the long-term) That's orthogonal to the simulation case, and doesn't extend much insight there. Said AI is not an entity in a simulated world, it's a simulated intelligence (let's say) in the real world, with means to reach the real world like access to the internet, talking to people back and forth, and so on. The AI we make needs those in order to be useful to us, and those are the things "AI containment" checks. An entity in a simulated world (in case we're that) could just as well not have access to anything like the above. E.g. the simulated world normally would not have access to facilities in the "top level" world. Say we are in such a world, we surely don't know of any way to dial into the "top level" equivalent of the internet, for example, nor do we have any back and forth talk with the "top level" users. Assuming we're in a simulation, in our case, we developed our bodies and thought through evolution in the simulated world (one of many species), whereas our AI was explicitly made to produce human like thinking output, and was trained with our own thoughts and data. This automatically makes it more compatible to create arguments "in our language" that persuade us. We? At best we could put up some kind of signs that we don't know if anybody in the top level ever checks, plus we don't know their motives, modes of thinking, or anything like that - or even if it's anything close to ours to be influenced by our level of arguments. And communication (already doubtful) is our surest best - for a program in a simulated environment can't really materialize as a physical actor in the top level we need to add too many assumptions and problems to solve. And they could always patch up any glitch as it appears, like we do in our code. Worse: the simulation code could be fully functional and pure, with nowhere the equivalent of a buffer overflow or other such means of escape. Worse: the simulated world could be at a totally different, much higher, and impossible to us, with our simulation-level intelligence to even comprehend - same way the simplistic code of an NPC character in a game has no bearing to us.