And what is that movement? What can you tell us about that?

Yeah, so the starting point of our research was the obvious observation that LLMs in particular are notoriously bad at saying, I don't know, at recognizing the limit of their own knowledge.

So we wondered whether Interacting with AI, with LLMs, would interfere with our capacity to say, I don't know.

So we conducted a series of experiments.

So the very first experiment to understand a little bit whether there is an effect, we designed six difficult questions.

And LLMs typically hallucinate because these are answers about visual details in movies, for example.

So something where the training of LLMs is not probably included in the training set of the LLMs.

Could you explain to us a little bit what is happening? What have you observed and the paper that you wrote? Because I think that to me, this is one of the most fascinating things that I discovered this year.

What made you want to investigate the impact of that on us, the impact of AI on our willingness to say, I don't know?

Yeah, exactly.

So on the one hand, to say, I don't know, is a very important human capability.

On the other hand, AI is notoriously reluctant to say, I don't know.

It has an answer to virtually every question.

So we wondered what happens to this important human capability when interacting with a system that doesn't have that capability.

So we designed a series of studies.

So what are the risks here? Are we a risk of being know-it-alls, Valerio? Like, am I at risk of being someone who just never admits I'm wrong?

What made you want to investigate the impact of that on us, the impact of AI on our willingness to say, I don't know?

Yeah, exactly.

So on the one hand, to say, I don't know, is a very important human capability.

On the other hand, AI is notoriously reluctant to say, I don't know.

It has an answer to virtually every question.

So we wondered what happens to this important human capability when interacting with a system that doesn't have that capability.

So we designed a series of studies.

So what are the risks here? Are we a risk of being know-it-alls, Valerio? Like, am I a risk of being someone who just never admits I'm wrong?