Dr. Arthur Lee

So faced with these choices, the participants made decisions where sometimes they made honest recommendations to the other client, but sometimes if the financial kickback was big enough, out of their own selfish reasons, they would say, hey, I think this one is better for you.

So they lie to get more money in the task.

That is, we have brain images of when people are lying, and we have brain images of when people are not lying.

And using machine learning, we can try to build a decoder that can classify, based on looking at somebody's brain, is this brain more likely to be somebody lying or more likely to be somebody not lying?

Just a very simple comparison test.

And we showed that if we just build a simple predictor like that, a statistical model, we can predict when people are lying or not pretty well about 70% of the time.

So that was the initial setup of the problem where if we approach this as like a lie detection problem, it may seem like we can detect lying pretty well based on brain activities.

So that was phase one.

So in the second phase, what we did was, okay, now we have a purported lie detector.

Let's see how much of a lie detector it actually is.

Let's see if it actually measures lying, if it says it is a lie detector.

So this is where it gets into the validity component.

So we devised a second task where participants played some similar game.

But in this game, they never told the other client which option is better for them.

They just said, I would prefer you choose this option or I would prefer you choose that option.

So there was technically no lying involved.

You didn't have to say this option is better for you or that option is better for you.

The strange thing was that if we take the lie decoder, the lie prediction model from the first study, and apply it in this dataset where nobody's lying, they're just being selfish or not selfish on channels, it actually predicts when somebody is being selfish versus not being selfish well above chance, almost similar to like the rate at which we detect lying.

So this was our demonstration that, look, we built a lie detector based on a simple premise of there being lie trials and non-lie trials, but if we apply it in a context where there's no lying involved, it predicts this other category of selfish versus non-selfish behaviors.

This is showing that