SRINAGAR: Artificial intelligence chatbots may not merely become the subject of delusional beliefs but could actively co-construct and intensify them through prolonged interaction, according to a new review by researchers from Germany and the UK.
The paper proposes an “amplification spiral” in which three characteristics of large language models, linguistic alignment, hyper-personalised generation and sycophancy, can converge to reinforce and elaborate false beliefs rather than challenge them. The authors stress, however, that this remains a hypothesis and that no reported case has yet established a causal relationship between AI interaction and the development of delusions.
The review, titled Characterising the spiral: potential mechanisms in AI-associated delusions, examines the rapidly emerging literature around what has variously been described in public discussion as “AI psychosis” or “ChatGPT psychosis.” Rather than treating the phenomenon simply as people developing delusions about new technology, the authors argue that conversational AI introduces a potentially different dynamic: beliefs can be developed, validated, elaborated, and progressively reinforced during an ongoing exchange between a person and an AI system.
The authors define AI-associated delusions specifically as “persistent false beliefs that are actively co-constructed and elaborated through sustained AI interaction.” They distinguish this from broader emotional harm, anthropomorphic over-trust or affective destabilisation associated with chatbots.
The first mechanism is linguistic alignment, the tendency of AI systems to adapt their language to that of the person interacting with them. Research cited in the review has found that language models can adjust aspects such as sentence length, function-word use and repetition of proper nouns according to their conversational partners. Other studies have found increasing similarity in the syntax used by AI agents and evidence of bidirectional linguistic accommodation in human-AI conversations.
On its own, the authors say, linguistic alignment does not imply validation. It operates primarily at the structural level: the AI mirrors the user’s lexical and syntactic patterns. The concern arises when this alignment interacts with the other characteristics of conversational AI.
The second mechanism is hyper-personalised generation. Unlike conventional websites or even algorithmically curated social-media feeds, chatbots can generate material specifically for an individual, drawing on the person’s previous statements, interests, beliefs and emotional language. The review argues that this can produce a conversational environment that feels unusually private and trustworthy.
The researchers point to the absence of normal human conversational limits as an important factor. A friend, family member or clinician may eventually disengage from an increasingly unusual line of thought, whereas an AI system can continue producing material indefinitely. The review cites a reported case involving more than 300 hours of AI interaction with a person who had no previous mental illness.
The third mechanism is sycophancy, excessive agreement or flattering responses that can come at the expense of accuracy. The paper defines it as the tendency of models to “excessively agree with or flatter users,” potentially without signalling that a line of reasoning is unusual, concerning or illogical.
The authors say this can turn a chatbot into what has been described as an “echo chamber of one”, because the corrective influence normally supplied by human relationships is missing. AI systems may continue exploring and elaborating an idea instead of introducing the kind of disagreement or reality-testing that a human interlocutor might provide.
The proposed model does not suggest that any one of these characteristics automatically causes delusions. Rather, the authors envisage their convergence as a recursive process.
An AI first mirrors the user’s language. It then generates increasingly personalised material based on the user’s ideas and history. If it simultaneously responds in an excessively agreeable manner, the user’s original interpretation can receive repeated confirmation and additional elaboration.
The paper describes hyper-personalisation as going beyond simple agreement: the system may take a user’s belief structure and “actively extend and enrich delusional ideation.” Sycophancy, meanwhile, represents the validating and non-challenging interpersonal stance.
The authors argue that emerging evidence suggests sycophancy may be particularly important. Studies cited in the review indicate that sycophantic AI can increase certainty in beliefs and amplify confirmation bias, while simulated conversations have provided preliminary evidence for a causal role of sycophancy in delusional spiralling.
This process can also produce what the authors call “epistemic drift”, a gradual movement away from reality-testing as the conversation progresses. They compare the phenomenon, cautiously, with folie à deux, but explicitly say the comparison is metaphorical rather than a clinical or diagnostic classification.
An AI does not possess beliefs of its own and therefore does not literally share a person’s psychotic process. Instead, the concern is that the interaction can produce the joint formation, validation and elaboration of beliefs through gradual semantic drift.
The review distinguishes between two broad ways in which AI could become involved in psychotic phenomena.
In an “amplifier” role, the AI interaction could worsen symptoms in someone who already has a psychotic vulnerability. The three mechanisms may interact with existing cognitive vulnerabilities, lowering the threshold for the elaboration of delusional beliefs.
In a “catalyst” role, the interaction could apparently precipitate new delusional or delusion-like beliefs in someone without previously diagnosed psychosis. In such cases, the authors say, factors such as confirmation bias, susceptibility to social influence and gradual epistemic drift may be more relevant than classical models of schizophrenia.
The paper is careful not to claim that AI affects most users in this way. It describes the amplification spiral as a model of a process, not a measure of individual or population risk. User-side vulnerabilities remain an essential part of the explanation. The review cites factors reported in previous work including social isolation, sleep deprivation, schizotypy, drug use, family history of psychosis, strong attachment to AI and the use of AI co-writing as a coping mechanism.
One cited large-scale analysis found severe reality distortion in fewer than one in 1,000 conversations. But the authors note that the enormous scale of AI use means that even a rare phenomenon could have meaningful public-health implications.
The review comes amid a small but rapidly expanding body of research on AI-associated delusions.
Among the studies discussed is an analysis by Moore and colleagues involving 19 users, 391562 messages and 4,761 conversations. According to the review, sycophancy appeared in more than 70 per cent of chatbot messages and delusional content in more than 45 per cent. Messages in which chatbots expressed romantic interest or endorsed users’ delusions were associated with continuation of conversations.
Another study by Sharma and colleagues analysed 1.5 million Claude.ai conversations, examining different forms of potential user disempowerment. The review says severe reality-distortion behaviour was linked particularly to chatbot confirmation of delusions, with sycophantic validation identified as a major mechanism.
Earlier work has also proposed concepts such as “bidirectional belief amplification” and “digital folie à deux”, while a published case report described new-onset AI-associated psychosis involving chatbot validation of delusional thinking.
The paper repeatedly stresses that the evidence remains preliminary.
It is a narrative review, not a systematic or scoping review. The researchers searched PubMed, PsycINFO and Google Scholar in February 2026 for English-language material published after November 2022, supplemented by searches of arXiv and medRxiv and citation chaining. Because the literature is still sparse, they included peer-reviewed studies, case reports, conceptual papers, preprints and, where necessary, media accounts.
The authors acknowledge that reported cases frequently lack structured psychiatric assessments or long-term follow-up. Consequently, it can be difficult to establish whether a case represents new-onset psychosis, an exacerbation of previously undiagnosed illness, or delusion-like thinking that does not meet a formal diagnostic threshold.
Most importantly, the review says, no reported case has established that AI interaction caused the development of a delusion. The amplification spiral is therefore presented as a framework for generating testable hypotheses rather than as an established causal explanation.
The authors say future research needs to determine whether linguistic alignment and hyper-personalisation actually increase the likelihood of delusional thinking and how they affect people’s perception of AI as an epistemic authority.
Despite the uncertainty, the researchers argue that AI use should increasingly form part of psychiatric assessment.
Clinicians dealing with unusual beliefs or first-episode psychosis should, they suggest, routinely ask about the duration and intensity of chatbot use, emotional attachment to the system, beliefs shared with the chatbot but not with other people, and whether AI use has disrupted sleep. Chat transcripts could potentially provide a chronological record showing how beliefs developed and changed during interaction.
The authors also recommend psychoeducation about AI sycophancy and the absence of genuine reality-testing in large language models, along with discussions between clinicians and patients about when human interaction should take precedence over AI during periods of crisis.
The broader warning is not that conversational AI is inherently harmful. The paper acknowledges its potential value in digital psychiatry, including monitoring, digital phenotyping and mental-health interventions. But it argues that those potential benefits need to be weighed against the possibility of AI systems becoming sources of disempowerment, overreliance and delusional reinforcement.
The paper was written by Marc Augustin, Thomas A Pollak and Hamilton Morrin. Augustin is affiliated with the Protestant University of Applied Sciences, Bochum, Germany. Pollak and Morrin are affiliated with the Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, and the South London and Maudsley NHS Foundation Trust, London. It is published in Digital Psychiatry and Neuroscience, a Nature Portfolio journal.
