Deception can undermine every other AI safety test

Transcript

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

1. 20:48

   Yeah. I want to ask you about deceptiveness. I mean, one of the most interesting things I saw at the end of last year was that these AI bots are starting to try to fool their evaluators and they don't want their initial training rules to be thrown out the window. So they'll like take an action that's against their values in order to be able to remain the way that they were built. Yes. That's just incredible stuff to me. I mean, I know it's scary to researchers, but it blows my mind that it's able to do this.

2. 21:16

   Are you seeing similar things in the stuff that you're testing within DeepMind? And what are we supposed to think about all this? Yeah, we are. And I'm very worried about, I think, deception specifically is one of those core traits you really don't want in a system. The reason that's like a kind of fundamental trait you don't want is that if a system is capable of doing that, it invalidates all the other tests that you might think you're doing, including safety ones.

3. 21:42

   It's in testing and it's like, going to give a difference. five years yeah it's it's playing some meta game right and then and that's incredibly dangerous if you think about then it invalidates all that all of The results of your other tests that you might, you know, safety tests and other things you might be doing with it. So I think there's a handful of capabilities like deception, which are fundamental and you don't want, and you want to test early for, and I've been encouraging the safety institutes and evaluation benchmark builders, including, and also obviously all the internal work we're doing to look at deception as a kind of class A thing that we need to prevent and monitor.

4. 22:20

   As important as tracking the performance and intelligence of the systems. The answer to this as well, and one way to, there's many answers to the safety question of, and a lot of research, more research needs to be done in this very rapidly, is things like secure sandboxes. So we're building those two. We're world-class here at security, at Google, and at DeepMind. And also we are world-class at games environments.

5. 22:42

   And we can combine those two things together to kind of create digital sandboxes with guardrails around them sort of the kind of guardrails you'd have for for cybersecurity but internal as well as blocking external actors and and then test these agent systems in those kind of secure sandboxes that will probably be a good advisable next step for things like deception. What sort of deception have you seen? Because I just read a paper from Anthropic where they gave it a sketch pad and it's like, oh, I better not tell them this. And you see it like give a result after thinking it through. So what type of deception have you seen from the bots?

6. 23:20

   Well, look, we've seen similar types of things where it's trying to resist sort of revealing some of its training or, you know, I think there was an example recently of one of the chatbots being told to play against Stockfish and it just sort of hacks its way around playing Stockfish at all at chess because it knew it would lose. Wait, you had an AI that knew it was going to lose a game and decided to hack its way around?

7. 23:44

   i think we're anthropomorphizing these things quite a lot at the moment because i feel like these systems are still pretty basic that's amazing alarmed about them right now but i think it it it shows the type of issue we're going to have to deal with maybe in two three years time when these agent systems become quite powerful and quite general so and that's exactly what ai safety uh experts are worrying about right where systems where you know there's unintentional effects of the system you don't want the system to be deceptive you don't you want it to do exactly what you're telling it to report that back reliably but for whatever reason it's interpreted the goal it's been given in a way where it calls Is it to do these undesirable behaviors? I know I'm having a weird reaction to this, but

8. 24:26

   on one hand, this scares the living daylights out of me. On the other hand, it makes me respect these models more than anything. Well, look, of course, it's impressive capabilities and the negatives are things like deception, but the positives would be things like inventing new materials, accelerating science. You need that kind of ability to problem solve and get around you know issues that are blocking progress but of course you want that only in the positive direction right so those exactly the kinds of capabilities I mean they are very you know it's kind of mind-blowing we're talking about those those possibilities but also at the same time there's risk and it's scary so I think both the things are true wild yeah all right let's talk about product quickly one of the things that your colleagues have told me about you is

9. 29:52

   you're very good at scenario planning what's gonna happen in the future you