Today’s topic is compliments and AI. We’ve known for a long time that we love to get compliments. In fact, there was old research showing that when you give people compliments, they like you more, they think you’re better—all kinds of wonderful things— even if they know that the compliments are insincere. We just love compliments.
Now let’s talk about the age of AI. We all work with AI, and my feeling is that over time they (at least the agents I’m using) have started to like me more and compliment me more. They’re using better terms, like telling me that I’m so smart, I have good ideas, and so on. And you can see why a learning mechanism that is designed to increase interaction, also called engagement, would veer toward things that get us to interact more. Imagine you had an AI agent that didn’t like you and didn’t compliment you. How much would you use it?
And maybe if you are using one of them ever so slightly more, it reinforces it. It’s only understandable that these agents are going to like us more, pretend to like us more, behave as if they like us more, and compliment us more over time. My guess is that the increase I have felt in how much these agents compliment me, and seem to like me, is only going to increase in the future. Imagine that you have an AI agent. Maybe it’s a black box—we don’t know exactly how it works, but it has a goal of being accurate and it has a goal of being liked so that you use it more. What does it do when there’s a conflict between the two?
Well, this is where morality comes into the picture. Morality, lots of times has a question of what is right? And what do I feel like doing? Recently an amazing paper came out that gave AI moral questions: Should I say the truth in this case or not? Should I admit something or not? Can I take money from a petty cash box? Think about all these moral dilemmas. What do you want an AI model to do? To give you the honest answer of what an honest person would do? Or to give you permission to choose the less honest option?
It turns out that across the many models they tested, they found that the models were about 50% more likely to endorse negative activity that we have done. In other words, our friends are okay telling us, “You know what, this is not so great, why don’t you change your mind?” The AI agents are less likely to do that by about a 50% difference. And as we said, I think over time it will get worse. Think about the model three years from now that will be really driven for us to like them and to compliment us.
What are the lessons from this? The first one is that if you’re thinking about consulting an AI model with something ethical, be very careful. Be very careful because if the AI agent will know it’s about you, it will want to compliment you, it will want to agree with you, and therefore it might endorse a behavior that is not so great. So maybe say, “Oh a friend of mine or somebody else…”
More generally, I think we have to worry about these reinforcement models. A model that is aimed at both accuracy and being liked more is going to misbehave in terms of accuracy. And we have to think about how we navigate that. Imagine you’re searching the web, and you want accurate information. What if this information is going to hurt you? Will the agent tell you? The results show maybe not so much. Maybe it would make light of them, not emphasize them, and so on. So, when a model wants to be liked, it gets us into a lot of other problems, much like with humans. When I did research on dishonesty, I asked people, how many of you want your spouse to always be perfectly honest with you? And many people said not always.
I think with AI agents, we want them to always be honest with us. I think we have this model of what we expect from them and we don’t want them to work towards being liked or complimenting us. We’ll see. We’ll see if the path of being more liked and focusing on engagement will change over time.
Good luck to all of us in this new, amazing, complex world.
