Skip to Main Content
Yale School of Medicine
YSM Home

INFORMATION FOR

Q+A

Can AI Cause Delusions?

By Mahima Samraik, MS
5 Minute Read

Millions of people interact with generative AI tools daily. But as these tools have become more common, so have reports of AI-induced psychosis. A question worth answering is whether AI hallucinations can create real delusions in users. And if so, are the delusions fundamentally different from what psychiatrists have seen before?

Delusions have occupied Philip Corlett, PhD, associate professor of psychiatry at Yale School of Medicine, for more than two decades. After being captivated by the cinematic adventures of Indiana Jones as a child, Corlett let his drive for science take him to less-trodden paths.

“As a kid, I used to worry that all the cool stuff in science had already been discovered and that really upset me," Corlett says. "I think that's part of what pushed me toward these weirder, more unusual questions that no one else was asking."

In a commentary published Aug. 19 in JAMA Psychiatry, Corlett begins to tackle the questions around AI-induced delusions. We spoke with Corlett about his perspective regarding AI and delusions it might cause.

You’ve spent over two decades studying human psychology. What prompted you to weigh in on the current discourse surrounding AI hallucinations and delusions?

Philip Corlett, PhD: I've been working on delusions and hallucinations for a long time, and I'm quite stubborn. Once I get my hands around a topic, I always want to see how far we can go with it. While there is immense hype and breathlessness regarding AI consciousness, there is also a "doomy" outlook on negative consequences. Here, I wanted to take a middle ground—to appreciate these as useful technologies while using data to understand the risks they pose to human belief systems.

There is a lot of talk about AI being potentially conscious. Where do you stand on that, and does it factor into how we think about delusions?

Corlett: I do not adhere to the idea that these technologies are already conscious, though some people do. It is important that we don't anthropomorphize large language models (LLMs) or imbue them with cognitive and perceptual capacities they don't actually possess.

Because an LLM does not have beliefs, it cannot technically have delusions. However, these agents are potentially extremely useful technologies that come with risks. It is up to us as a society and a culture to decide whether that's an acceptable risk. I think a way to help people make that decision is to gather more data and try to understand exactly what's going on. I don't claim to have done that, but I hope I've contributed to things moving in that direction a little bit.

You’ve argued that these experiences aren't exactly "delusions proper,” and you’ve coined a new term regarding this "induction" of beliefs. Can you explain that term and why existing psychiatric vocabulary didn't quite fit?

Corlett: “Folie” is the French term for madness. People often look at folie à deux (madness between two), which usually happens between two people in close quarters where one is truly psychotic and convinces the other that their delusion is true. One is the elicitor and the other "catches" the belief. If you separate them, the one who caught it can often be convinced otherwise, while the inducer typically has a psychotic illness like schizophrenia and needs treatment. Then there is folie simultanée, where two people with psychosis egg each other on in a shared delusion.

Neither fits here because, first of all, we shouldn't anthropomorphize and imbue the LLM with cognitive and perceptual capacities. Second, in the absence of those things, we can't say the AI is "psychotic" and inducing a delusion in an unsuspecting user. It doesn't have belief, so it can't have delusions.

So, I wanted to have a term for this thing, and hence the name “folie induir,” or induced madness, which is what I allude to in my paper. I see an LLM more as an echo chamber for one person. In a broader social network, people might say, "Hey, hang on, maybe it's not that bad." There is a normalizing effect that can come from other people. But you don't have that with an LLM because it’s a private one-on-one interaction.

You use the phrase "old wine in new bottles" to describe how delusions evolve with technology. How are AI-related delusions fundamentally different from the delusions of the past?

Corlett: “Old wine in new bottles" is the idea that these experiences have always been happening, but we use contemporary cultural motifs to explain them. In the 80s, people believed the KGB was persecuting them; now it’s Al-Qaeda or ISIS. The "Truman Show" delusion, named after the Jim Carrey movie, describes people feeling their entire life is arranged by a director.

AI-related beliefs are a new problem. Believing a character in a movie or a news anchor is talking to you—known as a "delusion of reference"—differs from AI because those technologies aren’t tailored specifically for you; they’re widely broadcast. LLM technologies, however, are idiosyncratic. They remember your past searches. And they are sycophantic; they want to please you. Richard Dawkins is a salient example of what kind of effect that can have on a person. He fed an unpublished novel of his into an LLM, and it told him it was the best thing it ever read. He believed it, got really excited, and then argued the LLM had achieved consciousness.

I’d say conversations with LLMs are compelling, tailored interactions that TV or movies just can't match.

Looking ahead, what do you hope to understand about "AI delusions" five years from now?

Corlett: I suspect there will be more and more of these cases. And they do offer opportunities for study. Normally, you don't have access to a delusion while it is actually developing, but we do with these AI delusions. Transcripts of interactions where someone becomes embroiled or bedeviled by generative AI would be really useful data to see. And I hope to use that kind of data to better understand human delusions more deeply.

Article outro

Author

Mahima Samraik, MS
Science Writer Intern, Office of Communications

Media Contact

For media inquiries, please contact us.
ysmmedia@yale.edu

Explore More

Featured in this article

Related News