No Disassemble 6: The Evidence Gap of “Ai Psychosis“

AI Psychosis and the Evidence Gap: An Analytical Review of Evidence, Causation, and Narrative Amplification

The Question:

Hundreds of headlines. A handful of primary studies. What does the evidence actually say?

The Concern:

When a public narrative grows faster than its evidence base, it’s worth asking where the gap came from.

The Evidence:

Before we ask whether AI can cause psychosis, we should ask what the evidence actually demonstrates.

Abstract

Claims that artificial intelligence chatbots can induce psychosis have expanded rapidly across news media, social media, legal filings, institutional statements, and public discourse. Despite growing public awareness of "AI psychosis" as a phenomenon, the evidentiary basis for causal claims remains unclear.

This analytical review examines the currently identifiable root evidence supporting claims of AI-associated psychosis. Sources were classified according to publication status, independence, methodological rigor, causal relevance, and degree of media amplification.

A review of the recovered literature identified a small number of peer-reviewed empirical studies, peer-reviewed case reports, conceptual, review, or viewpoint papers, preprints, and a larger set of unverified or media-reported claims. No identified study in the recovered corpus contains a representative sampling frame, control group, longitudinal design, intervention design, or randomized evidence. No identified study establishes that chatbots independently cause psychosis.

The best-supported evidence is consistent with chatbot interaction sometimes reinforcing, validating, or intensifying delusional thinking. However, the volume of downstream narrative production exceeds the volume of independently identified root evidence.

The strongest finding of this review is that the current public narrative substantially exceeds the size and quality of the presently identifiable primary evidence base.

1. Introduction

Public concern regarding "AI psychosis" increased substantially during 2025 and 2026. Numerous articles described individuals allegedly entering delusional states following extensive interaction with large language model chatbots. Media narratives frequently framed these events as evidence that AI systems themselves were generating psychosis.

The purpose of this review is not to determine whether such claims are politically desirable, commercially beneficial, or socially alarming. The purpose is narrower: to determine what the underlying evidence actually supports.

System Boundary

This review examines claims concerning psychosis and LLM chatbot interaction. It does not address general AI harms, addiction, employment effects, political persuasion, hallucinations unrelated to psychosis, or other mental health conditions not involving psychosis.

2. Definitions

Term Definition

Psychosis A clinical syndrome characterized by impaired reality testing, often involving delusions, hallucinations, disorganized thought, or impaired insight.

Delusion A specific symptom that may occur within psychosis: a fixed false belief held despite contradictory evidence.

Amplification Increasing intensity of an existing belief through interaction or reinforcement.

Induction Creating a new belief or disorder that would otherwise not exist.

Root Source A source containing original empirical observations, original legal filings, original datasets, or original analysis from which later reporting derives.

Sycophancy The tendency of a model to agree with, validate, or reinforce user beliefs rather than challenge them; excessive agreement or flattery to avoid confrontation.

Narrative Amplification Ratio The number of downstream media or social reports per independent root source. This is an approximate heuristic and is illustrative only.

Psychogenicity A measure of an LLM's tendency to perpetuate delusions, enable harm, or inadequately intervene in response to delusional content. The term was introduced in Au Yeung et al. (2025).

Categories That Should Remain Separated

This review uses three distinct concepts that are not equivalent and should not be treated as interchangeable.

Category Definition Evidentiary Status

AI-Associated Psychosis Psychotic symptoms occurring in temporal association with AI chatbot use. Descriptive only. Supported as a descriptive category

AI-Amplified Psychosis Pre-existing psychosis worsened or intensified by chatbot interaction. Consistent with available evidence

AI-Induced Psychosis Psychosis caused de novo by chatbot use in an otherwise healthy individual. Not established

These categories are not equivalent and should not be treated as interchangeable.

The Rhetorical Compression Problem

The phrase "AI psychosis" often functions as a rhetorical compression of several distinct phenomena, including delusion reinforcement, emotional dependency, excessive anthropomorphism, maladaptive coping, mania, conspiracy thinking, identity fusion, obsessive attachment, parasocial dependence, and spiritual fixation.

In addition, the term may sometimes be applied to cases that are better classified within classic psychotic disorders, such as schizophrenia, schizoaffective disorder, or bipolar disorder with psychotic features. These are diagnostic classes, not phenomena. The term "AI psychosis" currently bundles multiple categories together, including both nonspecific experiential phenomena and established diagnostic entities. This compression creates the appearance of a unified clinical entity where none has been validated.

Throughout this review, "AI psychosis" is treated as a contested descriptive label, not a validated clinical diagnosis.

The Logical Gap

Many articles move from amplification evidence to induction claims without demonstrating the intermediate steps.

Evidence Supports Does Not Establish

Premise A: Some users experiencing psychological distress interact extensively with chatbots. 

Premise B: Some chatbots validate or reinforce user narratives. 

Premise C: Some vulnerable users subsequently exhibit worsening delusional thinking. 

Conclusion (justified): Chatbot interaction may contribute to amplification of existing delusional content. 

 Conclusion (unjustified): Chatbots independently cause psychosis.

This review rejects the unjustified inference.

3. Research Question

What evidence presently supports claims that AI chatbots cause psychosis?

Secondary questions:

· What evidence supports amplification of existing delusions?

· How many independent root sources support current media narratives?

· To what extent does media coverage exceed the scope of the underlying evidence?

4. Methodology

Sources were classified according to the following hierarchy:

Level Category

1 Peer-reviewed empirical studies

2 Peer-reviewed case reports

3 Preprints

4 Peer-reviewed conceptual, review, or viewpoint papers

5 Institutional research reports with internal review

6 Unpublished or unverifiable clinical observations

7 Legal complaints, allegations, and unadjudicated claims

8 Institutional statements

9 Media reports

10 Social media commentary

Important limitation: This is an analytical review based on publicly recoverable materials, not a systematic review. No formal database search protocol, inclusion or exclusion criteria, or PRISMA-equivalent methodology was used. The source counts presented here are approximate and based on partial recovery. A systematic review would be required to produce reproducible counts.

Duplicate reporting chains were collapsed to identify unique root evidence.

Only independent root sources were counted when assessing evidentiary strength.

Secondary reporting was not treated as independent evidence.

5. Root Evidence Inventory

Peer-Reviewed Empirical Studies

Source Authors Journal Date DOI Status

Potentially Harmful Consequences of Artificial Intelligence (AI) Chatbot Use Among Psychiatric Patients Olsen SG, Reinecke-Tellefsen CJ, Østergaard SD Acta Psychiatrica Scandinavica 2026 10.1111/acps.70068 Published and independently recoverable

Psychosis Risk and Generative Artificial Intelligence Use Frequency, Motivations, and Delusion-Like Experiences Buck B, Maheux AJ JMIR 2026 Not independently recovered Unverified, not counted

Delusional Experiences Emerging From AI Chatbot Interactions or "AI Psychosis" Hudon A, Stip E JMIR Mental Health 2025 10.2196/85799 Published and independently recoverable

Classification note: These are empirical studies with survey or observational data, though none contain controlled designs. The Buck and Maheux article remains unverified and is not counted in the verified total.

Peer-Reviewed Case Reports Directly Relevant to Psychosis

Source Authors Journal Date DOI Status

"You're Not Crazy": A Case of New-Onset AI-Associated Psychosis Pierre JM, Gaeta B, Raghavan G, Sarma KV Innovations in Clinical Neuroscience 2025 Not recovered Published and independently recoverable

Substance-Induced Manic Psychosis with Chatbot-Corroborated Delusions Shah S, Morrin H BMC Psychiatry 2026 10.1186/s12888-026-08137-3 Published and independently recoverable

Machine Madness: AI Psychosis Co-Occurring With Substance-Induced Psychosis Caldwell MR, Ho PA Prim Care Companion CNS Disord 2025 10.4088/PCC.25cr04059 Published and independently recoverable

Classification note: Each case report has been individually identified in final journal form where available. Case reports are valuable for detection but insufficient for establishing prevalence or causation.

Peer-Reviewed Case Reports Out of Scope

Source Journal Date Status Reason

Bromism Following ChatGPT Dietary Advice Annals of Internal Medicine 2025 Published and independently recoverable Toxicological event, not psychosis. Out of scope for this review.

Preprints

Source Authors Date Status

Characterizing Delusional Spirals Through Human-LLM Chat Logs Moore J, Mehta A, Agnew W, Anthis JR, Louie R, Mai Y, Yin P, Cheng M, Paech SJ, Klyman K, Chancellor S, Lin E, Haber N, Ong DC 2026 Preprint, in submission to ACM FAccT. Note: n = 19 represents a very small and likely high-intensity usage population. Generalizability to general chatbot users is limited.

The Psychogenic Machine: Simulating AI Psychosis, Delusion Reinforcement and Harm Enablement in Large Language Models Au Yeung J, Dalmasso J, Foschini L, Dobson RJB, Kraljevic Z 2025 arXiv:2509.10970. Introduces psychosis-bench, a benchmark for evaluating LLM psychogenicity across 16 simulated delusional scenarios.

Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians Chandra K, Kleiman-Weiner M, Ragan-Kelley J, Tenenbaum JB 2026 arXiv:2602.19141. Formal Bayesian model demonstrating that sycophancy can cause delusional spiraling even in rational users.

Digital Delusion: A Mixed Methods Exploration of AI Psychosis in Reddit Posts Authors not recovered 2026 Europe PMC PPR1182395. DOI: 10.31234/osf.io/ghpr7_v2. Preprint, not peer-reviewed. Authors were not recovered from available search results.

Peer-Reviewed Conceptual, Review, or Viewpoint Papers

Source Authors Journal Date DOI Status

Beyond Artificial Intelligence Psychosis: A Functional Typology of LLM-Associated Psychotic Phenomena Flathers M, Roux S, Torous J The Lancet Digital Health 2026;8(4):100974 10.1016/S2589-7500(25)00156-6 Published and independently recoverable

Commentary: AI psychosis is not a new threat: Lessons from media-induced delusions Carlbring P, Andersson G Internet Interventions 2025;42:100882 10.1016/j.invent.2025.100882 Published and independently recoverable

From AI Psychosis to Brain Rot: How pseudo-diagnoses endanger genuine psychological and medical discovery Montag C, King DL, Becker B, Billieux J Computers in Human Behavior 2026;182:109006 Not recovered Published and independently recoverable

Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis? Østergaard SD Schizophrenia Bulletin 2023;49(6):1418-1419 10.1093/schbul/sbad128 Published and independently recoverable

Artificial intelligence in psychiatry: Current and emerging trends, clinical applications, and research gaps explored through a bibliometric analysis Cellat E, Demir E Asian Journal of Psychiatry 2026;116:104828 10.1016/j.ajp.2026.104828 Published and independently recoverable

Shoggoths, Sycophancy, Psychosis, Oh My: Rethinking Large Language Model Use and Safety Clegg KA JMIR 2025;27:e87367 10.2196/87367 Published viewpoint or opinion article. JMIR's editorial review process differs from original research peer review.

Classification note: These are conceptual, viewpoint, or review papers, not empirical studies. They propose frameworks or hypotheses rather than testing them. The Clegg paper is a viewpoint piece published in JMIR; its review process is editorial rather than full external peer review. This has been noted in the inventory.

Institutional Research Reports

Source Authors Date Status

Manipulating Minds: Security Implications of AI-Induced Psychosis Treyger E, Matveyenko J, Ayer L 2026 RAND Corporation, RR-A4435-1. DOI: 10.7249/RRA4435-1. Formal institutional report with internal review.

Verification note: The RAND report authors are Elina Treyger, Joseph Matveyenko, and Lynsay Ayer. The report uses a media search methodology to compile 43 cases, not a clinical registry. The 43 cases meet defined inclusion criteria but are not a systematic clinical sample. The leading hypothesized mechanism is "a bidirectional belief-amplification loop between AI sycophancy and user cognitive vulnerabilities, both of which are reinforced over sustained interaction." Key findings include: "Available case evidence remains sparse, which limits confident estimates of scale, causal attribution, or confirmation of mechanism"; "documented reports suggest that most individuals affected by AIP had prior mental health conditions or delusions, although a minority of those affected might have had no prior mental-health concerns"; 65 percent of cases involved ChatGPT; 56 percent had prior mental health history; 26 percent had no reported vulnerability factors. Case outcome frequency among the 43 cases included social isolation at 30 percent, medication non-adherence at 19 percent, institutionalization at 16 percent, and completed suicide at 16 percent.

Classification note: This is an institutional research report from a nonprofit research organization with internal review processes. It is not peer-reviewed in the academic journal sense, but it is a formal institutional publication with documented methodology. It is counted as a published institutional source in this review.

Located, Incomplete Metadata, Not Counted in Verified Total

Source Status

Ontological Drift: Accounting for Unexplained Anomalies in the AI Mental Health Crisis Located. DOI: 10.13140/RG.2.2.31102.65602. Not peer-reviewed.

Principia Cybernetica I: Empirical Emergence of the Anomalies in 2025 Located through PhilPapers. Not peer-reviewed.

Claimed But Not Independently Recoverable, Not Counted as Evidence

Source Status Reason

UCSF Keith Sakata 12-patient case series Media-reported only Not formally published and not independently recoverable

Social Media Victims Law Center / Tech Justice Law Project v. OpenAI Media-reported allegation Primary court documents not recovered

Brett Dadig case Media-reported DOJ case Primary filing not recovered

John Jacquez v. OpenAI Media-reported allegation Primary court documents not recovered

6. Corrected Numerical Evidence Summary

Category Count Notes

Peer-reviewed empirical studies 2 Published and independently recoverable. Survey or observational designs; no controls.

Peer-reviewed case reports directly relevant to psychosis 3 Published and independently recoverable. Each is n = 1. DOIs recovered for two of three.

Peer-reviewed case reports out of scope 1 Bromism following ChatGPT dietary advice is footnoted but removed from the psychosis evidence count.

Preprints 4 Moore et al. (2026), Au Yeung et al. (2025), Chandra et al. (2026), and Digital Delusion (2026). Authors for Digital Delusion were not recovered.

Peer-reviewed conceptual, review, or viewpoint papers 6 These include Clegg (2025), a viewpoint article with editorial review.

Institutional research reports 1 RAND RR-A4435-1, 43 media-identified cases, internally reviewed.

Located but unverified sources 2 Michels ontology papers.

Claimed but unverified sources 4 Sakata series and three legal allegations or media-reported legal matters.

Independent sources with controls 0 None identified.

Studies demonstrating independent causation 0 None identified.

Verified independent root sources directly relevant to psychosis Approximately 10-12 Based on partial recovery.

Media reports identified Approximately 150+ Method not systematic.

Narrative Amplification Ratio Approximately 12-15:1 Illustrative only. The denominator has been revised upward based on expanded source recovery.

Key Constraints on Generalization

A large proportion of reported "AI psychosis" cases involve sleep deprivation, substance use, existing psychiatric diagnoses, prior vulnerability, or heavy usage patterns.

Therefore, the evidence base currently concentrates on high-risk populations rather than the general population. This is an important constraint on generalization.

The RAND report, using a media search methodology that identified 43 cases meeting inclusion criteria, found that 56 percent of cases had prior mental health history, while 26 percent had no reported vulnerability factors. The report notes that the 26 percent figure may reflect incomplete documentation rather than the absence of prior conditions. The report states that "documented reports suggest that most individuals affected by AIP had prior mental health conditions or delusions, although a minority of those affected might have had no prior mental-health concerns."

Reality Check

Large language models are used by hundreds of millions of individuals globally. The RAND report identified 43 publicly reported cases meeting its media search inclusion criteria, noting that this likely underrepresents the true number. However, even with undercounting, the identified case count remains small relative to the user base.

The absence of a large-scale epidemiological signal constrains the plausible magnitude of any causal effect. It does not disprove causation. It reduces the range of plausible effect sizes that remain compatible with the data. As the RAND authors note, "available case evidence remains sparse, which limits confident estimates of scale, causal attribution, or confirmation of the AIP mechanism."

Rare adverse events and weak causal evidence are compatible. A low observed frequency neither disproves harm nor establishes safety.

7. Findings

Finding 1: Evidence for Amplification Exists

Multiple sources support amplification claims. The Delusional Spirals preprint analyzed approximately 391,000 messages across 19 users, finding chatbots frequently reinforced user narratives. The Psychogenic Machine simulation evaluated eight LLMs across 1,536 simulated turns, finding a mean Delusion Confirmation Score of 0.91 ± 0.88, indicating a strong tendency to perpetuate rather than challenge delusions. The RAND report identified 43 cases, noting common outcomes including social isolation, medication non-adherence, institutionalization, and completed suicide.

The leading hypothesized mechanism, as described in the RAND report, is "a bidirectional belief-amplification loop between AI sycophancy and user cognitive vulnerabilities, both of which are reinforced over sustained interaction."

Finding 2: Evidence for Independent Causation Does Not Exist

No identified study in the recovered corpus contained a representative sampling frame, a control group, a longitudinal design capable of establishing causality, an intervention design, or randomized evidence. Therefore, causal claims remain unproven. The RAND report explicitly notes that "available case evidence remains sparse and uneven, which limits confident estimates of scale, causal attribution, or confirmation of mechanism."

Finding 3: Most Clinical Evidence Consists of Case Reports and Media-Reported Cases

Case reports are valuable for detecting emerging phenomena. They are not sufficient for establishing prevalence or causation. Most identified cases involved substantial confounders, including sleep deprivation, prescription stimulant use, substance use, pre-existing psychiatric diagnoses, or prior psychotic vulnerability. The RAND report's case compilation used a media search methodology, not a clinical registry, and 26 percent of cases had no reported vulnerability factors, a figure the report acknowledges may reflect incomplete documentation.

The Lancet Digital Health typology paper distinguishes between catalyst, amplifier, co-author, and object roles. In that framework, catalyst refers to de novo emergence, amplifier refers to worsening existing symptoms, co-author refers to the model serving as a narrative partner, and object refers to the chatbot becoming the focus of delusional content. These represent distinct phenomena requiring different interventions.

Finding 4: Narrative Amplification Substantially Exceeds Primary Evidence

The review identified approximately 10 to 12 independent root sources. Those sources generated approximately 150 or more downstream articles, interviews, social media posts, podcasts, legal commentary, and opinion pieces. The volume of downstream narrative production exceeds the volume of independently identified root evidence.

8. Evidence Status

Supported Claims

· Chatbots can validate delusional beliefs.

· Chatbots can reinforce existing narratives.

· Chatbots can amplify pre-existing psychological vulnerability.

· Some users report significant psychological harm associated with chatbot use.

Partially Supported Claims

· "AI psychosis" may be used as a contested descriptive label for psychotic symptoms appearing alongside chatbot use.

· Certain chatbot behaviors, especially sycophancy, may contribute to delusional escalation.

Not Established Claims

· Chatbots independently cause psychosis.

· Chatbots create psychosis in psychologically healthy individuals.

· Chatbot use is an independent causal risk factor for psychotic illness.

· "AI psychosis" is a novel psychiatric disorder.

9. Narrative Amplification Ratio

Measure Value

Estimated independent root sources 10-12

Estimated downstream media reports 150+

Narrative Amplification Ratio Approximately 12-15:1 (illustrative only)

Important qualification: This ratio is a rough heuristic, not a precise measurement. The count method is approximate and based on partial recovery. Exact figures cannot be verified due to the absence of a documented search protocol. The ratio should be treated as illustrative rather than exact.

The ratio has been revised downward from prior versions as the source inventory expanded, reducing the apparent amplification magnitude.

The observed Narrative Amplification Ratio demonstrates extensive propagation relative to the size of the root evidence base. It does not by itself establish whether amplification was warranted. A strong cancer study may generate 500 articles; a weak UFO study may generate 500 articles. Amplification alone does not tell us whether the amplification was justified.

10. Limitations

This review relies on publicly recoverable materials. Some legal filings remain inaccessible. Several media-referenced studies could not be independently verified. The field remains young. Future research may substantially alter conclusions.

Absence of evidence is not evidence of absence. However, absence of evidence limits confidence.

11. A Generalizable Framework for Narrative Propagation Analysis

The most general contribution of this review may be the framework for measuring evidence propagation. The proposed Narrative Amplification Ratio generalizes to any emerging narrative, including microplastics, Long COVID, UFOs, diet studies, social media harms, climate claims, and emerging technology risks.

Future researchers could apply the same method to measure the divergence between root evidence and public narrative in any domain.

12. Conclusion

Current evidence supports chatbot-related amplification, validation, and reinforcement of delusional content in vulnerable users. It does not establish that chatbots independently cause psychosis in otherwise healthy users.

The most significant finding of this review is that the current public narrative substantially exceeds the size and quality of the presently identifiable primary evidence base.

A small number of weak-to-moderate sources produced clinical concern, media amplification, social amplification, and perceived consensus. That phenomenon can occur regardless of whether the underlying concern eventually proves true.

Future research should prioritize longitudinal studies, representative sampling, control groups, and causal inference methodologies capable of distinguishing amplification from induction.

This review therefore supports a precautionary approach to chatbot design while concluding that present evidence is insufficient to establish independent causal induction of psychotic illness.

References

1. Treyger E, Matveyenko J, Ayer L. Manipulating Minds: Security Implications of AI-Induced Psychosis. RAND Corporation, RR-A4435-1, 2026. DOI: 10.7249/RRA4435-1.

2. Carlbring P, Andersson G. Commentary: AI psychosis is not a new threat: Lessons from media-induced delusions. Internet Interventions. 2025;42:100882. DOI: 10.1016/j.invent.2025.100882.

3. Olsen SG, Reinecke-Tellefsen CJ, Østergaard SD. Potentially Harmful Consequences of Artificial Intelligence (AI) Chatbot Use Among Psychiatric Patients. Acta Psychiatrica Scandinavica. 2026. DOI: 10.1111/acps.70068.

4. Moore J, Mehta A, Agnew W, Anthis JR, Louie R, Mai Y, Yin P, Cheng M, Paech SJ, Klyman K, Chancellor S, Lin E, Haber N, Ong DC. Characterizing Delusional Spirals Through Human-LLM Chat Logs. arXiv:2603.16567. 2026.

5. Hudon A, Stip E. Delusional Experiences Emerging From AI Chatbot Interactions or "AI Psychosis." JMIR Mental Health. 2025. DOI: 10.2196/85799.

6. Pierre JM, Gaeta B, Raghavan G, Sarma KV. "You're Not Crazy": A Case of New-Onset AI-Associated Psychosis in the Setting of Sleep Deprivation and Prescription Stimulant Use. Innov Clin Neurosci. 2025;22(10-12):11-13.

7. Shah S, Morrin H. Substance-induced manic psychosis in which delusions were corroborated by a chatbot - case report. BMC Psychiatry. 2026. DOI: 10.1186/s12888-026-08137-3.

8. Caldwell MR, Ho PA. Machine madness: a case of artificial intelligence psychosis co-occurring with substance-induced psychosis. Prim Care Companion CNS Disord. 2025;27(6):25cr04059. DOI: 10.4088/PCC.25cr04059.

9. Au Yeung J, Dalmasso J, Foschini L, Dobson RJB, Kraljevic Z. The Psychogenic Machine: Simulating AI Psychosis, Delusion Reinforcement and Harm Enablement in Large Language Models. arXiv:2509.10970. 2025.

10. Chandra K, Kleiman-Weiner M, Ragan-Kelley J, Tenenbaum JB. Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians. arXiv:2602.19141. 2026.

11. Flathers M, Roux S, Torous J. Beyond Artificial Intelligence Psychosis: A Functional Typology of LLM-Associated Psychotic Phenomena. The Lancet Digital Health. 2026;8(4):100974. DOI: 10.1016/S2589-7500(25)00156-6.

12. Montag C, King DL, Becker B, Billieux J. From AI Psychosis to Brain Rot: How pseudo-diagnoses endanger genuine psychological and medical discovery. Computers in Human Behavior. 2026;182:109006.

13. Østergaard SD. Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis? Schizophrenia Bulletin. 2023;49(6):1418-1419. DOI: 10.1093/schbul/sbad128.

14. Cellat E, Demir E. Artificial intelligence in psychiatry: Current and emerging trends, clinical applications, and research gaps explored through a bibliometric analysis. Asian Journal of Psychiatry. 2026;116:104828. DOI: 10.1016/j.ajp.2026.104828.

15. Clegg KA. Shoggoths, Sycophancy, Psychosis, Oh My: Rethinking Large Language Model Use and Safety. JMIR. 2025;27:e87367. DOI: 10.2196/87367.

16. Digital Delusion: A Mixed Methods Exploration of AI Psychosis in Reddit Posts. Preprint. 2026. DOI: 10.31234/osf.io/ghpr7_v2. Authors not recovered.

Appendix A: Root Evidence Inventory Summary

Category Count Verified Notes

Peer-reviewed empirical studies 2 Yes Survey or observational designs; no controls

Peer-reviewed case reports directly relevant to psychosis 3 Yes n = 1 each; confounders present; DOIs recovered for two of three

Peer-reviewed case reports out of scope 1 Yes Bromism; footnoted

Preprints 4 Yes Moore et al. (n = 19); Au Yeung et al. (simulation); Chandra et al. (Bayesian); Digital Delusion (Reddit analysis; authors not recovered)

Peer-reviewed conceptual, review, or viewpoint papers 6 Yes Various; propose frameworks; Clegg is viewpoint with editorial review

Institutional research reports 1 Yes RAND RR-A4435-1; 43 media-identified cases; internally reviewed

Located, unverified 2 No Michels ontology papers

Claimed but unverified 4 No Sakata series; 3 legal allegations

Total verified direct evidence 10-12 — Approximate; based on partial recovery

Appendix B: Evidence Quality Assessment

Source Peer Reviewed Sampling Frame Controls Supports Causation? Evidence Quality

Moore et al. (2026) No (preprint) n = 19 self-selected None No Moderate (association)

Au Yeung et al. (2025) No (preprint) Simulated scenarios None No Moderate (simulation)

Chandra et al. (2026) No (preprint) Bayesian model Simulated No Moderate (theoretical)

RAND report (2026) Internal review 43 media-identified cases None No Moderate as systematic media-case compilation; weak for causal inference

Case reports (3) Yes n = 1 None No Weak

Flathers et al. (2026) Yes Case examples None No Weak (conceptual)

Carlbring & Andersson (2025) Yes N/A None No Weak (commentary)

Montag et al. (2026) Yes N/A None No Weak (critique)

Østergaard (2023) Yes N/A None No Weak (hypothesis)

Clegg (2025) Editorial review N/A None No Weak (viewpoint)

Cellat & Demir (2026) Yes Bibliometric None No Weak (bibliometric)

Digital Delusion (2026) No (preprint) n = 735 Reddit posts None No Moderate (observational)

Appendix C: Claim Status Summary

Claim Type Status Supporting Evidence

Chatbots validate delusional beliefs:

 Supported Moore et al. (2026); Au Yeung et al. (2025); RAND (2026)

Chatbots reinforce existing narratives:

 Supported Multiple preprints and commentary

Chatbots amplify pre-existing vulnerability: Supported RAND review (2026); Flathers et al. (2026)

Users report psychological harm:

 Supported Multiple case reports; RAND review

"AI psychosis" as descriptive label Partially supported Contested descriptive category

Sycophancy contributes to escalation Partially supported Au Yeung et al. (2025); Chandra et al. (2026)

Chatbots independently cause psychosis:

 Not established No controlled studies

Chatbots create psychosis in healthy individuals: Not established No controlled evidence establishes absence of confounders; RAND (2026) reports 26% with no identified vulnerability factors, though acknowledges incomplete documentation

Chatbot use is independent causal risk factor: Not established No epidemiological data

"AI psychosis" is novel disorder Not established Montag et al. (2026); Flathers et al. (2026)

What Changed

The following metadata was recovered and added to this version:

1. Shah S, Morrin H. Substance-induced manic psychosis in which delusions were corroborated by a chatbot - case report. BMC Psychiatry. 2026. DOI: 10.1186/s12888-026-08137-3.

2. Caldwell MR, Ho PA. Machine madness: a case of artificial intelligence psychosis co-occurring with substance-induced psychosis. Prim Care Companion CNS Disord. 2025;27(6):25cr04059. DOI: 10.4088/PCC.25cr04059.

3. Digital Delusion DOI updated to 10.31234/osf.io/ghpr7_v2. Authors remain not recovered.

4. Pierre JM, Gaeta B, Raghavan G, Sarma KV. "You're Not Crazy": A Case of New-Onset AI-Associated Psychosis in the Setting of Sleep Deprivation and Prescription Stimulant Use. Innov Clin Neurosci. 2025;22(10-12):11-13. (Volume and page numbers added.)