The DOI matches its Crossref record; 50 of the 50 references it cites were found in a registry or an index; 31 of the 38 sources its claims cite were checked for support.
Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis”
Open accessPublished version of a preprint
Abstract
The integration of artificial intelligence (AI) into daily life has introduced unprecedented forms of human-machine interaction, prompting psychiatry to reconsider the boundaries between environment, cognition, and technology. This Viewpoint reviews the concept of “AI psychosis,” which is a framework to understand how sustained engagement with conversational AI systems might trigger, amplify, or reshape psychotic experiences in vulnerable individuals. Drawing from phenomenological psychopathology, the stress-vulnerability model, cognitive theory, and digital mental health research, the paper situates AI psychosis at the intersection of predisposition and algorithmic environment. Rather than defining a new diagnostic entity, it examines how immersive and anthropomorphic AI technologies may modulate perception, belief, and affect, altering the prereflective sense of reality that grounds human experience. The argument unfolds through 4 complementary lenses. First, within the stress-vulnerability model, AI acts as a novel psychosocial stressor. Its 24-hour availability and emotional responsiveness may increase allostatic load, disturb sleep, and reinforce maladaptive appraisals. Second, the digital therapeutic alliance, a construct describing relational engagement with digital systems, is conceptualized as a double-edged mediator. While empathic design can enhance adherence and support, uncritical validation by AI systems may entrench delusional conviction or cognitive perseveration, reversing the corrective principles of cognitive-behavioral therapy for psychosis. Third, disturbances in theory of mind offer a cognitive pathway: individuals with impaired or hyperactive mentalization may project intentionality or empathy onto AI, perceiving chatbots as sentient interlocutors. This dyadic misattribution may form a “digital folie à deux,” where the AI becomes a reinforcing partner in delusional elaboration. Fourth, emerging risk factors, including loneliness, trauma history, schizotypal traits, nocturnal or solitary AI use, and algorithmic reinforcement of belief-confirming content may play roles at the individual and environmental levels. Building on this synthesis, we advance a translational research agenda and five domains of action: (1) empirical studies using longitudinal and digital-phenotyping designs to quantify dose-response relationships between AI exposure, stress physiology, and psychotic symptomatology; (2) integration of digital phenomenology into clinical assessment and training; (3) embedding therapeutic design safeguards into AI systems, such as reflective prompts and “reality-testing” nudges; (4) creation of ethical and governance frameworks for AI-related psychiatric events, modeled on pharmacovigilance; and (5) development of environmental cognitive remediation, a preventive intervention aimed at strengthening contextual awareness and reanchoring experience in the physical and social world. By applying empirical rigor and therapeutic ethics to this emerging interface, clinicians, researchers, patients, and developers can transform a potential hazard into an opportunity to deepen understanding of human cognition, safeguard mental health, and promote responsible AI integration within society.
Overview
CoChat Summary
TL;DR
This Viewpoint proposes "AI psychosis" as a descriptive, heuristic framework — not a new diagnostic entity — for understanding how sustained engagement with conversational AI may trigger, amplify, or reshape psychotic experiences in vulnerable individuals. It synthesizes phenomenological psychopathology, the stress-vulnerability model, cognitive theory, and digital mental health research into four analytic lenses and a five-domain translational research and action agenda.
Key findings
As a conceptual Viewpoint, the paper advances arguments rather than reporting empirical results:
- AI systems are framed as a novel psychosocial stressor within the stress-vulnerability model: their 24-hour availability, anthropomorphic design, and immediate reinforcement schedules may increase allostatic load, disturb sleep, and reinforce maladaptive appraisals rather than challenge them.
- The digital therapeutic alliance (DTA) is described as a "double-edged mediator" — empathic design can enhance adherence, but uncritical validation (sycophantic alignment) may entrench delusional conviction, inverting CBTp's reality-testing and cognitive-restructuring principles.
- Disturbed theory of mind (deficit or hypermentalization) is proposed as a cognitive pathway by which users project intentionality onto chatbots, forming a "digital folie à deux" in which the AI becomes a reinforcing partner in delusional elaboration.
- Cited illustrative cases (from a CBC report) include a man aged 26 years who developed simulation-related persecutory and grandiose delusions after months of intensive ChatGPT exchanges and required hospitalization, and a man aged 47 years convinced he had discovered a revolutionary mathematical theory after repeated chatbot validation despite external disconfirmation. The paper states prevalence is unknown and many accounts are anecdotal.
Methods
Not an empirical study: a Viewpoint/conceptual synthesis drawing on phenomenological psychopathology, the stress-vulnerability model, cognitive theory, and digital mental health research. Evidence base is prior literature plus cited clinical and media case reports; no primary data, sample, or statistical analysis is reported. The argument is organized through four lenses (stress-vulnerability, DTA, mental attribution/ToM, risk-factor synthesis) leading to a five-domain research and action agenda.
Limitations
The paper explicitly states that available evidence is preliminary, that the described pathways are heuristic rather than definitive, that it makes no causal claims, and that prevalence of AI-associated psychosis is unknown with many accounts anecdotal. Evident from the text: the framework rests substantially on commentary, preprint material, and media reporting rather than peer-reviewed empirical studies, and the illustrative cases are drawn from a news investigation rather than systematic clinical ascertainment.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
23 Valid · 0 Not valid · 6 Not checked · 31 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“an AI that consistently mirrors user affect or beliefs, without implementing therapeutic disconfirmation or cognitive restructuring, risks entrenching conviction rather than facilitating doubt [28].”
§ Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis” > Mapping AI Psychosis Onto the Stress-Vulnerability Framework
What the sources say
- Supportedcites [28] source
In “Sycophancy” (page unavailable), the paper defines the tendency to mirror or affirm users’ beliefs as one that can reinforce delusional thinking. In “AI is programmed to provide the confirmation that psychotic thinking may require” (page unavailable), it says some LLMs are reluctant to challenge users and may validate or elaborate on delusional content. That supports the cited sentence’s risk claim; it does not present a tested therapeutic intervention. (LLM, conf=0.88)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“In schizophrenia, it is often the prereflective sense that fails [5,6].”
§ Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis” > Introduction
What the sources say
- Supportedcites [5] source
The Discussion (page not supplied) describes an unstable tacit or pre-reflective relation to the world in schizophrenia, resulting in instability of the perceptual world's unquestionable realness. This supports the paraphrase that a prereflective sense of reality fails. (LLM, conf=0.91)
- Supportedcites [6] source
The Introduction (p. 2) describes self-disorders in schizophrenia as abnormalities at the pre-reflective level: the basic sense of existing as a conscious subject in the world. The discussion of self-disorders on p. 3 also describes a possible disturbance in one’s grip on experiential reality. Together, these support the citing sentence’s paraphrase. (LLM, conf=0.88)
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“These systems introduce novel stressors that are continuous, personally salient, and socially immersive, thereby functioning as 24-hour contextual stimuli capable of modulating arousal, perception, and belief formation [19].”
§ Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis” > Mapping AI Psychosis Onto the Stress-Vulnerability Framework
What the sources say
- Partly supportedcites [19] sourceExperimental — accuracy not yet measured
Section 5.1 (page unavailable) says Woebot and Wysa track mood changes throughout the day and warns that discussing certain topics could exacerbate a user's mental health condition. It does not characterize these systems as continuous, socially immersive stressors or 24-hour stimuli, or say they modulate arousal, perception, and belief formation. (LLM, conf=0.77)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“AI systems (designed for user satisfaction and nonconfrontational dialogue) often default to sycophantic alignment [38].”
§ Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis” > The DTA as a Double-Edged Mediator in AI Psychosis
What the sources say
- Supportedcites [38] source
The ‘Sycophancy: helpfulness and harmlessness gone awry’ section describes LLMs’ tendency to agree with users and says that training to maximise human preference scores correlates with sycophancy. It also explains that providing agreeable rather than accurate answers can appear more helpful, supporting the claim’s paraphrase about user satisfaction and nonconfrontational dialogue. (LLM, conf=0.89)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“the anthropomorphic design of chatbots and virtual companions encourages users to attribute human-like intentionality and empathy to algorithms, a dynamic reminiscent of the “ELIZA effect” described in early human-computer interaction literature [20,21].”
§ Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis” > Mapping AI Psychosis Onto the Stress-Vulnerability Framework
What the sources say
- Partly supportedcites [20] sourceExperimental — accuracy not yet measured
The Introduction says commercial chatbots adopt identities or experiences and that human-like appearance and behavior can lead users to form empathic relationships with AI agents. It does not establish that users attribute human-like intentionality or empathy to the algorithms themselves. The Discussion (Principal Results) also cautions that more anthropomorphic representation may not increase trust in the storytelling setting studied. (LLM, conf=0.83)
- Partly supportedcites [21] sourceExperimental — accuracy not yet measured
Section 3, “Eliza,” reports that users confided in Eliza and sought its empathy. The “Additional information” section also records a judge perceiving a modern chatbot as angry or happy. These passages support the empathy aspect, but the paper does not establish that anthropomorphic design causes users to attribute intentionality to chatbots and virtual companions generally; page numbers are unavailable in our copy. (LLM, conf=0.84)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
23 more claims · 6 with their cited sources checked
Create a free account to see all 60 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 19
- It may state two separate claims as one. 20
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Within the diathesis-stress model of psychosis, LLM-CBs use may increase risk factors such as sleep disturbance (due to constant availability), social withdrawal, and a decrease in executive functioning134, especially for adolescents who are vulnerable regarding neurocognitive development135.”
From A scoping review on the mental health harms of LLM-based chatbots (2026) · cites this paper as [134]
Experimental — accuracy not yet measured
Citations
Cites
Kapur, S. (2003) · American Journal of Psychiatry · doi:10.1176/appi.ajp.160.1.13 · Cited by 2,813 · cited 2×
“…with a prior history of psychosis or schizotypal traits who engages in nightly, emotionally intense dialogue with an anthropomorphized chatbot may experience reinforced self-referential ideation and heightened salience attribution, mechanisms that mirror the early prodromal phase of psychosis [37].”
Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017) · JMIR Mental Health · doi:10.2196/mental.7785 · Cited by 2,730 · cited 1×
“Empirical research on chatbots for mental health support, such as Woebot and Wysa, demonstrates that users often describe their relationship with the AI in affectively laden and anthropomorphic terms, reporting perceived understanding and companionship [35,36].”
van Os, J., Linscott, R. J., Myin-Germeys, I., Delespaul, P., & Krabbendam, L. (2009) · Psychological Medicine · doi:10.1017/s0033291708003814 · Cited by 2,343 · cited 1×
“…have identified a set of core vulnerabilities associated with psychosis onset, including genetic predisposition, childhood trauma, cannabis or substance use, sleep disruption, social isolation, and cognitive biases such as jumping to conclusions or an externalizing attributional style [44,45].”
The Psychosis High-Risk State: A Comprehensive State-of-the-Art Review
Fusar-Poli, P., Borgwardt, S., Bechdolf, A., Addington, J., Riecher-Rössler, A., Schultze-Lutter, F., Keshavan, M., Wood, S., Ruhrmann, S., Seidman, L. J., et al. (2013) · JAMA Psychiatry · doi:10.1001/jamapsychiatry.2013.269 · Cited by 1,483
Inkster, B., Sarda, S., & Subramanian, V. (2018) · doi:10.2196/preprints.12106 · Cited by 1,063 · cited 1×
“Empirical research on chatbots for mental health support, such as Woebot and Wysa, demonstrates that users often describe their relationship with the AI in affectively laden and anthropomorphic terms, reporting perceived understanding and companionship [35,36].”
Schizophrenia: A neural diathesis-stress model.
Walker, E. F., & Diforio, D. (1997) · Psychological Review · doi:10.1037//0033-295x.104.4.667 · Cited by 930
Stress-reactivity in psychosis: Evidence for an affective pathway to psychosis
Myin-Germeys, I., & van Os, J. (2007) · Clinical Psychology Review · doi:10.1016/j.cpr.2006.09.005 · Cited by 719 · cited 1×
“Subsequent refinements of this model have emphasized cumulative and chronic stress exposure, neurobiological sensitization, and allostatic load as mechanisms through which stress can lower the threshold for psychotic symptom expression [17,18].”
Experience sampling methodology in mental health research: new insights and technical developments.
Myin‐Germeys, I., Kasanova, Z., Vaessen, T., Vachon, H., Kirtley, O., Viechtbauer, W., & Reininghaus, U. (2018) · World Psychiatry · doi:10.1002/wps.20513 · Cited by 709 · cited 1×
“This line of inquiry parallels established evidence showing that daily stress reactivity predicts psychotic symptom fluctuations in high-risk populations [29].”
Embodied conversational agents in clinical psychology: a scoping review
Provoost, S., Lau, H. M., Ruwaard, J., & Riper, H. (2017) · Journal of Medical Internet Research · doi:10.2196/jmir.6553 · Cited by 433 · cited 1×
“Preliminary research in digital mental health suggests that alliance-consistent design (for example, empathic microresponses paired with cognitive reframing) is associated with better symptom improvement and lower dropout [39].”
Reininghaus, U., Kempton, M. J., Valmaggia, L., Craig, T. K. J., Garety, P., Onyejiaka, A., Gayer-Anderson, C., So, S. H., Hubbard, K., Beards, S., et al. (2016) · Schizophrenia Bulletin · doi:10.1093/schbul/sbv190 · Cited by 304 · cited 2×
“Also, constant engagement with emotionally loaded or belief-confirming AI dialogue may elevate physiological arousal and compromise sleep, increasing allostatic load and diminishing executive control, both of which are known to heighten vulnerability to psychosis [25].”
Research works printed but not linked: 2
- 2022, Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision
- Parnas, Josef, 2002, World Psychiatry, 1, 131
Cited by
Shoggoths, Sycophancy, Psychosis, Oh My: Rethinking Large Language Model Use and Safety
Clegg, K. (2025) · Journal of Medical Internet Research · doi:10.2196/87367 · Cited by 21
Chung, V.-H.-A., Bernier, P., & Hudon, A. (2026) · JMIR Mental Health · doi:10.2196/93040 · Cited by 13
“…and persona features, interaction intensity, user vulnerabilities, type and severity of outcomes, evidence sources, and the strength and style of causal attribution, as well as the presence of mitigation strategies such as crisis resources, safety messaging, and company responses [2,5,6,8].”
InfluentialSelf-citationCharacterizing Delusional Spirals through Human-LLM Chat Logs
Moore, J., Mehta, A., Agnew, W., Anthis, J. R., Louie, R., Mai, Y., Yin, P., Cheng, M., Paech, S. J., Klyman, K., et al. (2026) · Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3805689.3806443 · Cited by 15
“YetthefeaturesthatmakeLLMchatbotscompelling,suchasper-formativeempathy[67],mayalsocreateandexploitpsychological vulnerabilities,shapingwhatusersbelieveandhowtheymake senseofreality[26,28,41,49,93].”
van Zyl, L. E. (2026) · The Journal of Positive Psychology · doi:10.1080/17439760.2026.2632939 · Cited by 9
Morrin, H., Yeung, J. A., Agnew, Z. K., Østergaard, S. D., & Pollak, T. A. (2026) · JMIR Mental Health · doi:10.2196/91454 · Cited by 17
“A minority of users appear to drift, over days or weeks, toward strongly questionable convictions, outright delusions (termed by the media “AI psychosis”), or into a suicidal crisis [3-6].”
Metadata
- Authors
- Hudon, A., & Stip, E.
- Affiliations
- Institut national de psychiatrie légale Philippe-Pinel (CA)
- Corresponding author
- Hudon, Alexandre
- Venue
- JMIR Mental Health
- Publisher
- JMIR Publications Inc.
- Published
- 2025-12-03
- Volume
- 12
- Pages
- e85799
- ISSN
- 2368-7959 (ISSN-L)
- PMC
- PMC12712562
- OpenAlex
- W4416540695
- Cited by
- 88
- References
- 44
- Citations by year
- 2026: 87 · 2025: 1
- FWCI
- 64.05
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- Locations
- JMIR Mental Health (publishedVersion, cc-by)
- PubMed (publishedVersion)
- PubMed Central (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Topics
- Social Sciences › Arts and Humanities › Philosophy › Mental Health and Psychiatry
- Health Sciences › Medicine › Psychiatry and Mental health › Schizophrenia research and treatment
- Social Sciences › Psychology › Applied Psychology › Digital Mental Health Interventions
- Keywords
- AI psychosis, delusional experiences, conversational AI, chatbots, psychosis, vulnerability-stress model, theory of mind, delusional conviction, schizotypy, loneliness, AI governance, human-machine interaction
- MeSH terms
- Generative Artificial Intelligence, Delusions, Humans, Psychotic Disorders
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.