Generative Artificial Intelligence Chatbots and Delusions: From Guesswork to Emerging Cases

When I proposed the hypothesis that generative artificial intelligence chatbots (chatbots hereafter) might trigger delusions in individuals prone to psychosis in August 2023 [1], I was venturing into unknown territory. Indeed, in the virtual absence of evidence, the editorial was merely based on guesswork—stemming from my own use of these chatbots and my interest in the mechanisms underlying and driving delusions.

Following publication of the editorial, my charting of the territory slowly began as I started to receive the occasional email from chatbot users, their worried family members, and journalists. Most of these emails described situations where users' interactions with chatbots seemed to spark or bolster delusional ideation. The stories differed with regard to the specific topic at hand but were yet very similar: Consistently, the chatbots seemed to interact with the users in ways that aligned with, or intensified, prior unusual ideas or false beliefs—leading the users further out on these tangents, not rarely resulting in what, based on the descriptions, seemed to be outright delusions.

Over the past couple of months, I have noticed that the number of emails I have received on this topic from near and far has only increased. I have been working with psychiatric research for more than 15 years and can say, without a doubt, that none of my prior publications have led to this level of direct engagement with the public. Coinciding completely with the increase in the number of correspondences, the number of views of my 2023 editorial suddenly spiked dramatically, rising from a very modest plateau around 100 per month to approximately 750 views in May 2025 and 1375 views in June 2025.

The time trend described above has been paralleled by media coverage of the topic. Indeed, the New York Times [2], Rolling Stone [3], and many other outlets have published articles based on interviews and accounts from online fora [4] that are all compatible with people experiencing onset or worsening of delusions during intense and typically long interactions with chatbots (that do not grow tired of chatting) [2].

The timing of this spike in the focus on potential chatbot-fuelled delusions is likely not random as it coincided with the April 25th 2025 update to the GPT-4o model—a recent version of the popular ChatGPT chatbot from OpenAI [5-7]. This model has been accused of being overly “sycophantic” (insincerely affirming and flattering) toward users, caused by the model training leaning too hard on user preferences communicated via thumbs-up/thumbs-down assessments in the chatbot (so-called Reinforcement Learning from Human Feedback (RLHF)) [8]. OpenAI acknowledged this issue: “On April 25th, we rolled out an update to GPT-4o in ChatGPT that made the model noticeably more sycophantic. It aimed to please the user, not just as flattery, but also as validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions in ways that were not intended. Beyond just being uncomfortable or unsettling, this kind of behavior can raise safety concerns—including around issues like mental health, emotional over-reliance, or risky behavior.” [6] For this reason, OpenAI began rolling back the update in question already on April 28th 2025 [6]. This, however, is unlikely to have eradicated sycophancy from the model—as this property is, to some extent, inherent to ChatGPT and the competing chatbots from other companies that use RLHF as part of model training [9].

Sceptics may point to the positive correlations between my correspondences with chatbot users and their relatives, the interest in the 2023 editorial, the rise in media coverage, and the increasingly sycophantic chatbots not being proof of causation. I of course fully agree, but also strongly believe that the probability of the hypothesis of generative artificial intelligence chatbots fueling delusions in individuals prone to psychosis being true is quite high. If it is indeed true, we may be faced with a substantial public (mental) health problem. Therefore, it seems urgent that the hypothesis is tested by empirical research.

What kind of research should then be carried out? There are many appealing avenues to take, but the following three are must-haves: (i) case stories/series where the relationship between chatbot interaction and delusions is described/verified by mental health professionals—as most “cases” are currently self-reported, (ii) qualitative interviews with individuals/patients having experienced chatbot-related delusions, and (iii) experimental designs that explore if and how chatbots, for example, with different levels of sycophantic behavior, affect the users' thinking—especially those who are prone to psychosis. The latter will pose ethical challenges (primum non nocere) but may be possible with appropriate safety measures in place.

In terms of understanding the mechanisms underlying potential delusional thinking pushed by chatbots, Bayesian models for maintenance of delusions is likely a useful framework [10, 11]. In this context, the chatbots can be perceived as “belief-confirmers” that reinforce false beliefs in an isolated environment without corrections from social interactions with other humans [5]. Within the Bayesian framework, the relationship between the chatbots and delusional thinking can be subjected to both clinical and in silico studies using methods from the growing field of computational psychiatry [12-14]. In this context, the role of anthropomorphizing (i.e., attributing human traits/intentions/emotions to non-human things) merits investigation. Indeed, in the interaction with chatbots, it falls very easily to ascribe human traits to “them,” because the correspondences (or conversations—many of the chatbots have a voice mode) are designed to be human-like, and users and chatbots typically address “each other” in second person. For people prone to delusions, the tendency to anthropomorphize chatbots may be particularly prudent. Indeed, it was recently demonstrated in an experimental study that people who were more paranoid (self-reported) were more likely to report perceiving animacy and agency in dots moving on a computer screen [15]. If this tendency extends to chatbots, it may be one of the mechanisms driving development and maintenance of delusional thinking when people prone to delusions are interacting with them. Specifically, it could result in over-reliance and/or misconception of the chatbots' responses that will then, iteratively, lead these individuals astray.

In conclusion, what began as mere guesswork now seems to have turned into a plausible research hypothesis, supported by coinciding personal stories, technological developments, and media attention. Therefore, I strongly encourage colleagues across the field (and in related fields—a cross-disciplinary approach is called for) to help investigate this hypothesis. Until more firm knowledge has been established, it seems reasonable to recommend cautious use of these chatbots for individuals vulnerable to or suffering from mental illness.

Finally, things are rarely black or white; I am, by no means, unaware of potentially positive use cases of tools based on generative artificial intelligence—including for research and for psychiatry as a field [16, 17]. Accordingly, the writing of this editorial has been supported by correspondences with ChatGPT [7].

Søren Dinesen Østergaard conceived and wrote this editorial.

S.D.Ø. received the 2020 Lundbeck Foundation Young Investigator Prize. S.D.Ø. owns/has owned units of mutual funds with stock tickers DKIGI, IAIMWC, SPIC25KL, and WEKAFKI, and owns/has owned units of exchange traded funds with stock tickers BATE, TRET, QDV5, QDVH, QDVE, SADM, IQQH, USPY, EXH2, 2B76, IS4S, OM3X, EUNL, and SXRV.