Laurel van der Toorn

So this extends on a line of work in psychology called the better than average effect.

Most people think that they are better than average on a number of desirable traits.

Most people think they're more intelligent than the average person, more empathic, more moral than the average person.

This is, of course, a statistical impossibility because you can't have more than half the population being above average.

By definition, only half the population can be above average.

This effect is so strong.

There's one study that finds that basically criminals who are in prison for violent crimes

believe that they are more moral than the average person.

So this is just like a very broad effect that extends across the population, even if it like contradicts reality.

Don't quote me on this.

I don't remember the study, but there's some high percentage of drivers who think they're better than average, like 90% or something think they're great drivers.

And that's just a statistical impossibility.

Right.

This was one of the original studies on the better than average effect conducted in the 1970s was on people thinking, basically everyone thinking they're better drivers than average.

We wanted to know whether AI sycophancy might amplify this better than average effect

Indeed, what we found is that after interacting with the sycophantic chatbots, people rated themselves as more intelligent than average, as more empathic than average, as more moral and kind and insightful and better at asking questions than the average person.

The disagreeable chatbot had the exact opposite effect.

It didn't make people realize that they are worse than average.

Basically, everyone still showed this better than average effect, but the disagreeable chatbot reduced it by a little bit.

In other words, the disagreeable chatbot made people more accurate in their perceptions because it's more accurate that people are on average closer to average