Iyad Rahwan with David Sloan Wilson - Evolving Alignment Podcast

Okay, great.

So that was quite a long gestation period.

That was published in 2019, but it really reflects a lot of work on your part.

Not only are you bringing in behavior, but you're bringing in ecology.

Behaviors are always enacted in environments, and the ecology of AI is really like the ecology of a multi-species communities with all the interactions and so on.

Let's bring in multi-level selection in the concept of major evolutionary transitions.

a topic dear to my own heart, but I'd love to hear from you as to why it's important to distinguish levels of selection and also this idea of major evolutionary transitions in which a unit becomes so highly cooperative that it becomes endowed with the properties of a single organism.

That's how cooperative it's become.

Yes, so maybe I should just build a bridge to this concept by saying that in my own research and in other people's research, we're working at different levels, right? So we're looking at different kinds of questions.

So I find myself often working on some proximate questions.

So for instance, we recently had a paper showing that when people delegate tasks to AI agents, they are more likely to cheat on those tasks.

So that's a very kind of proximate behavior that we were able to observe and quantify experimentally.

So if maybe you want to fire somebody or you want to maybe write a fake review of a product and you can delegate it to an AI, you're more willing to do that.

So now we diagnose this.

Now we can think about, well, why did this behavior evolve or emerge? Or perhaps the AI allows you this kind of affordance, doesn't push back, unlike your friend who might say, look, no, I'm not going to do this for you.

But then AI doesn't care.

So we start noticing differences.

And then you can start thinking, well, in what sort of ecology? Well, first of all, what happens if this becomes so prevalent? Everybody now has a kind of minions that do their dirty work for them.

So you can start projecting or speculating about like these more ecosystem level consequences of

something like this.

And then you can think,

well, what sort of ecology or what sort of environmental conditions should I create to prevent this kind of behavior? So, for example, perhaps we should audit delegation to AI agents and then impose consequences on them so that people feel a real kind of reputational consequences or even maybe financial consequences.

if they cheat using AI agents or they undertake unethical actions, for example, cheat on their taxes through AI.

So that's kind of like the background.

And the question now becomes, well, under what conditions can we make AIs more prosocial or more like moral or ethical or legal.