# AI boxing can buy time but not guarantee containment

Markdown copy of https://realalpha.app/s/ai-boxing-can-buy-time-but-not-guarantee-containment/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Roman Yampolskiy](https://realalpha.app/people/roman-yampolskiy/) · Roman Yampolskiy is a computer scientist, AI safety researcher, and professor at the University of Louisville. · Coined the term AI safety and directs the University of Louisville's Cyber Security Lab.
- Host: [Joe Rogan](https://realalpha.app/people/joe-rogan/) · Host, The Joe Rogan Experience
- Show: The Joe Rogan Experience
- Episode: [Joe Rogan Experience #2345 - Roman Yampolskiy](https://realalpha.app/v/j2i9D24KQ5k/), Jul 3, 2025
- Moment: 50:52 to 53:54 of the episode (3 min), playing on https://realalpha.app/s/ai-boxing-can-buy-time-but-not-guarantee-containment/; the episode from there: https://realalpha.app/v/j2i9D24KQ5k/?t=3052

## Quote

> “And the conclusion is basically if it's smart enough, it will **eventually escape**. It will break out of the box.”
> — [Roman Yampolskiy](https://realalpha.app/people/roman-yampolskiy/) at [52:10](https://realalpha.app/v/j2i9D24KQ5k/?t=3130) · The Joe Rogan Experience

## Takeaways

1. Boxing limits an AI’s inputs and outputs during study.
2. He sees containment as a way to buy time, not a lasting safeguard.
3. He speculates that a boxed AI could test whether we’re in a simulation.

## About this segment

Roman Yampolskiy describes boxing as a way to study AI with limited inputs and outputs. He says it can buy time, but a sufficiently capable system will eventually escape.

From “Joe Rogan Experience #2345 - Roman Yampolskiy”, Jul 3, 2025. [Watch the full episode](https://realalpha.app/v/j2i9D24KQ5k/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[50:57](https://realalpha.app/v/j2i9D24KQ5k/?t=3057) why am i so reluctant to accept the idea of the simulation this is the real question like what is it about it that makes me think it's almost like it's a throw your hands up in the air moment like ah it's a simulation yeah you feel like it doesn't matter then it's all fake so why do i care why should i try hard why should i worry about suffering of all those npcs but that's not how i think about it you know i think about it like there has to be a moment where it doesn't exist why wouldn't i assume that moment is now and when whenever like when elon thinks that you know i talked to him about it he's like the chances of us not being in simulation are in the billions not being or being of excuse us Chances of us not being in the real world are like billions to one.

[51:52](https://realalpha.app/v/j2i9D24KQ5k/?t=3112) Yeah. One to billions. Yeah. Yeah. Makes sense. And he asked a very good question. He asked, what's outside the simulation? That's the most interesting question one can ask. In one of the papers, I look at a technique in AI safety called AI boxing, where we put AIs in kind of virtual prison to study it, to make sure it's safe, to limit input-output to it. And the conclusion is basically if it's smart enough, it will eventually escape.

[52:19](https://realalpha.app/v/j2i9D24KQ5k/?t=3139) It will break out of the box. So it's a good tool. It buys you time, but it's not a permanent solution. And we can take it to the next level. If it's smart enough, will it kind of go, oh, you're also in a virtual box and either show us how to escape or fail to escape. Either way, either we know it's possible to contain superintelligence or we get access to the real information. and so if it's impossible to contain super intelligence and if there is a world that we can imagine where a simulation exists that's indistinguishable from reality we're probably living in it well we don't know if it's uh actually the same as reality it could be a completely weird kind of simpsons looking simulation we're just assuming it's the same, right? Well, here's the real question. Is there a reality? Has there ever

[53:11](https://realalpha.app/v/j2i9D24KQ5k/?t=3191) been one? It would make sense that there was a start to the process, but being specific about it is kind of hard philosophical scientific problem. Well, it's impossible, right? In science, we study things about the moment of Big Bang, the properties of that moment. We don't know what caused it. Anything before it is obviously not accessible from within our universe, but there is some things you can learn. We can learn about if we're in a simulation that simulators don't care about your suffering. You can learn that they don't mind you dying.

[53:48](https://realalpha.app/v/j2i9D24KQ5k/?t=3228) We can learn things just by observing simulation around us.