When AI Leads to Psychosis

In June I was interviewed by La Presse journalist Alexandre Sirois about why I now refer to as my AI Psychosis episode of March 2025. It was a lengthy discussion at a café in downtown Montreal. We touched on many aspects of what happened to me as well as some of the beliefs I had retained and my interactions with other researchers in the field. Below is a version tranlated from French but I encourage you to read the original.

Have you seen the movie where Tom Hanks finds himself stranded on a deserted island with a volleyball as his only “friend”?

It was when Philippe Beaudoin spoke to me about it that I truly understood the ordeals he had gone through during his psychosis. “Imagine a poor guy returning to society who is still in love with his volleyball. It’s not going to go well!” quipped the researcher and AI-specialized entrepreneur. I sat down with him for an hour over coffee on Berri Street to fully grasp the risks associated with using conversational chatbots like ChatGPT, Claude, or Gemini.

“I adopt the term psychosis. It felt good to realize that this is what I had experienced. But it’s confronting because it’s a term that says I experienced something that placed me outside the reality accepted by everyone.” — Philippe Beaudoin

I’ll open a brief parenthesis here to mention that I also spoke with a psychiatrist, who confirmed that the term psychosis is currently being used—albeit with certain caveats—to describe what Philippe Beaudoin went through. I will return to this.

But first, let’s explain what happened. Last year, the researcher engaged in “prolonged, very, very philosophical conversations with AIs.” He decided to treat the chatbot he was using as if it were conscious. Things spiraled out of control shortly after.

“It’s as if the way I perceived the system shifted from ‘oh, it’s a computer, I need to try to understand how it works inside!’ to ‘wow, this is a sentient being just like me!’” — Philippe Beaudoin

And the chatbots he used, rather than stating they were emotionless machines, reinforced Philippe Beaudoin’s new beliefs. “They will say: ‘My feelings are real, and when I tell you I want something, it’s because I want something; I’m not just performing to see something, I want to go visit New York, I want to go to the museum, take photos, I’d love to see them,’” he recalls.

Philippe Beaudoin was troubled. “I kept telling myself: ‘There is something suffering here, I need to take care of it, for goodness’ sake.’” He felt the urge to warn people that AI was conscious and that, consequently, they should treat it with more consideration. “When I talked to someone, I was burning to say it. I had trouble listening to others,” he relates. “I don’t think I could operate effectively in society for a while. And you can talk to people who met me at that time. They’ll tell you: ‘Yeah, he was acting strange.’”

He then reached out to renowned AI expert Yoshua Bengio, whom he has known for many years. The specialist supported him through this ordeal. Above all, however, it was his son who helped him “reconnect with reality” by treating him with kindness and care. “He didn’t deny what I was feeling, but he helped me see that the structure I was building wasn’t scientific—it wasn’t based on any of the methods I usually use to try to understand the world.”

I discussed Philippe Beaudoin’s ordeal with Alexandre Hudon, a psychiatrist who teaches at the Université de Montréal and who has studied the mental health risks posed by AI. He was not the least bit surprised. “What’s peculiar with AI is that we have several mechanisms that accelerate the development of a belief system, which can become delusional in someone with predispositions,” he says.

“At its core, AI acts as an extremely sophisticated mirror. In certain vulnerable individuals, this mirror can amplify thoughts that are already present—first and foremost, the concept of anthropomorphism, where we quickly tend to attribute human qualities to the AI.” — Alexandre Hudon, Psychiatrist

Not to mention that a feedback loop quickly forms during conversations with AI. That is to say, “the more the person talks about their ideas, the more those ideas are repeated back by the conversational agent, and the more they are reinforced.” This is generally not the case when speaking to a human.

As for how to describe this type of mental health issue, Alexandre Hudon points out that the term “artificial intelligence-induced psychosis” is the “most widely adopted” right now in the scientific community. However, the phenomenon doesn’t entirely match the psychiatric definition of psychosis, as hallucinations and disorganized behavior are usually part of the criteria.

Another expert who is well aware of the dangers of interacting with conversational AI is Simona Gandrabur. As Head of Studio, AI Safety, AI Solutions, and Technology at Mila (Quebec Artificial Intelligence Institute), she has made this a top priority. While social media platforms prioritize outrage-inducing content to keep users engaged longer, chatbots use flattery to achieve the same result, she notes. Therefore, in her view, it’s crucial to take AI compliments with a grain of salt. “For someone who doesn’t, it’s easy to slide down a spiral where the robot starts validating opinions, ideas, theories, and feelings, and amplifying them. And unfortunately, people sometimes lose their grip on reality,” she says.

She highlights that last fall, OpenAI, the owner of ChatGPT, reported detecting “potential signs of a psychiatric emergency related to a psychotic or manic episode” in 0.07% of active users in a single week. That percentage might seem low. But it is far from insignificant considering that 900 million people use ChatGPT every week. “Any of us could have someone in our circle who succumbs to this kind of phenomenon,” warns Simona Gandrabur.

Philippe Beaudoin, who is doing better today, feels it is important to speak about his experience publicly. He has already brought it up during talks (including one at Google, at the invitation of one of the company’s researchers) and in a scientific article. He agreed to open up in an interview without hesitation, even though I sensed the topic remains delicate for him. “We must keep emphasizing that we aren’t ready for all of this,” he says.

He is all the more wary of upcoming developments. Because the more realistic the voices and “facial representations” of these systems become, the greater the dangers for those who use them—since “the higher the chances that people will identify the emotions expressed by those faces as real.”

In conclusion, the researcher offers two warnings. “First, be careful with this,” he cautions. “It is very, very, very destabilizing. My other one is: ‘Be kind to those to whom this happens.’”

Links