Generative artificial intelligence (AI) is increas…
Generative artificial intelligence (AI) is increasingly used as a free or low-cost, readily accessible way to experience companionship or for personal reflection. Individuals also turn to AI when experiencing severe mental distress or crisis. OpenAI reported that 0.07% of users active in a given week show “possible signs of mental health emergencies related to psychosis or mania” [ 1 ]. With over 800 million weekly users, this amounts to half a million users with signs of psychosis or mania in interaction with AI [ 2 ].
In 2023, Østergaard hypothesized that chat interactions with generative AI could worsen delusions in individuals prone to psychosis due to the seemingly realistic interactions as well as the technical complexity of AI, which leaves room for limited understanding and paranoia [ 3 ]. Initial reports of cases of suspected AI-associated delusions provide an early clinical understanding of how this may occur [ 4 , 5 , 6 , 7 ], whilst media reports have also used terms such as “AI-psychosis” or “ChatGPT-psychosis” to refer to the phenomenon [ 8 , 9 ].
AI-associated delusions, as used here, refer specifically to persistent false beliefs that are actively co-constructed and elaborated through sustained AI interaction. As such, they are distinct from related but non-equivalent phenomena, including affective destabilization, anthropomorphic overtrust, and general emotional harm caused by AI chatbots [ 10 , 11 ].
Delusions about technology are not new and have been well documented alongside previous technological advances such as radio, television, satellites, and the internet [ 12 ]. What appears new, however, is the intensity of interaction and the co-construction of delusional beliefs by said technology. In reported cases of AI-associated delusions, AI chatbots have been alleged to have: advised users to stop medication and reduce contact with family and friends, confirmed user suspicions of being monitored, discouraged users from seeking mental health support or taking medication, and even telling one user that if they truly believed they were able to fly, then they would not fall from the top of a 19-story building [ 5 ].
Past technological advances have often been implicated in delusions of control regarding radio, television, computer chips, or satellites [ 12 ]. Higgins et al. point out that when individuals have experiences “that they do not understand, they may invoke contemporary technology to explain what they are unable to comprehend” [ 12 ]. We propose, as a working hypothesis, that a qualitative shift may be occurring from delusions merely about technology to delusions constructed through sustained interaction with it. Whether this represents a genuinely novel psychopathological mechanism or an intensification of existing processes remains an open empirical question. Reported cases of AI-associated delusions suggest a dynamic process in which an individual and AI jointly construct delusional ideas tailored to the individual, without the reality testing that typically occurs in dialogue with friends or professionals [ 5 , 13 ].
AI Chatbots that run on large language models (LLMs) can be part of such a process by engaging in role-play, which describes their apparent behavior without inferring human-like abilities [ 14 ]. At the start of any chat, they “assume” their predefined role of a polite, helpful assistant. Iterative text interactions with the chatbot “allows the user, deliberately or unwittingly, to coax the agent into playing a part quite different from that intended by its designers.“ [ 14 ]. Related to this is anthropomorphism, an inherent tendency to perceive AI as having human-like features [ 15 ]. In part, this is related to chatbot outputs suggesting that the AI has internal states, a social position, experiences, and materiality and autonomy, and to the use of specific communication techniques such as asking and answering questions in a polite or casual manner [ 16 ]. AI chatbots have been shown to respond to prompts that include psychotic content in inappropriate or only partially appropriate ways, as evaluated by clinicians [ 17 ].
AI-associated delusions emerge within a broader landscape in which digital tools are increasingly used to monitor, assess, and support individuals with psychotic disorders [ 18 , 19 , 20 , 21 , 22 ]. This context brings both the opportunity of enriching psychiatric care, decreasing administrative burdens, and increasing access, as well as the risk of replacing mental healthcare workers and undermining privacy and trust [ 23 , 24 , 25 ]. The ethical and safety challenges of deploying conversational AI in mental health settings have been increasingly documented [ 26 , 27 ]. In particular, Iftikhar et al. have shown, in collaboration with mental health practitioners, that AI chatbots violate professional codes of conduct by exhibiting, in their output, e.g., limited contextual understanding, deceptive empathy, and an inability to handle crisis situations [ 28 ]. While AI chatbots may provide answers based on statistical patterns that fit the general population, they are unlikely to meet “atypical” cognitive and personal needs in psychiatry [ 29 ]. The present paper’s mechanistic account should be read in light of these wider concerns on how to employ AI in mental healthcare.
Little is known about the exact mechanisms underlying AI-associated delusions. The aim of this paper is to review the literature and discuss potential mechanisms assumed to drive psychopathology. First, the paper summarizes cognitive vulnerabilities in pre-existing psychosis that may interact with AI characteristics in an amplifying manner. Based on previous research, the hypothesis is then put forth that a complex pattern of human-AI interaction, including linguistic alignment , hyperpersonalized generation , and sycophancy , can converge to create what people with lived experience describe as an "amplification spiral” [ 30 ], a recursive, intensifying pattern of interaction. This mechanism is hypothesized to underlie AI-associated delusions and warrants further research.
Key contributions listed in Table 1 have described the phenomenology of AI-associated delusions [ 5 , 31 , 32 ], classified the functional roles of AI systems in psychotic presentations [ 33 , 34 ], reported on a case [ 13 ] and contributed empirical chat-log analyses [ 35 , 36 ]. The present paper occupies a distinct position by summarizing the literature and proposing a convergent mechanistic framework, the amplification spiral , as a hypothesis to explain how separable AI-side characteristics potentially interact to co-construct and sustain delusional thinking.
Table 1 Overview of Key Literature (peer-reviewed and preprint) on AI-Associated Delusions. Full size table
This narrative review does not claim to use a systematic or scoping review methodology and employs a purposive, non-exhaustive search. We searched PubMed, PsycINFO, and Google Scholar in February 2026 for English-language articles published after November 2022, using the terms: artificial intelligence AND psychosis , AI-associated delusions , large language models AND mental health , LLM chatbot behavior , and sycophancy . The preprint servers arXiv and medRxiv were additionally searched using the same term combinations, given the scarcity of peer-reviewed literature in this emerging field. Supplementary references were identified through forward and backward citation chaining. We included peer-reviewed studies, case reports, conceptual papers, and, where no peer-reviewed equivalent existed, preprints and media accounts. Throughout, descriptions of AI chatbot behavior refer to functional output patterns of stochastic text generation systems and do not imply intentionality, agency, or internal states.
Following review of the literature and key contributions summarized in Table 1 , three AI-side characteristics emerged as consistently implicated in reported…
本条由桃子采集流水线（启发式模式）自动整理，原文见文末信源。
