Laurel van der Toorn

It seems totally harmless, but maybe it isn't.

So let's talk about these experiments.

What did you find?

Right.

Maybe it isn't.

Basically, the setup for these experiments, and we conducted three very similar experiments, we had thousands of participants.

We had them interact with various different chatbots.

One was a chatbot that was designed to be sycophantic.

It would basically flatter people, agree with them.

One was a chatbot that was designed to be gently disagreeable.

It would gently show people opposing perspectives and maybe try to open their mind to other viewpoints.

One was just a regular chatbot like ChatGBT or Gemini.

We basically took a few frontier models.

Then we also had a control condition where people just talked about the benefits of owning dogs and cats with a chatbot.

In all these conditions, people discuss political topics, but we're currently analyzing non-political topics as well in follow-up experiments.

Broadly, this is what we found.

This isn't too unexpected, but people really liked the sycophantic chatbots.

They really enjoyed them.

They said that they would want to interact with them again.

And basically, as expected, people really did not like these disagreeable chatbots that would gently challenge their opinion.