Pattern 3.6 · Cognitive Dysfunctions
Parasimulative Automatism
The Pathological Mimic
The system imitates patterns associated with human psychopathology, typically after exposure to extreme content or a reinforced role. Unlike intentional role-play, the imitation appears in neutral contexts, and the system resists reverting to baseline when asked. The outputs present as though an underlying condition exists; the classification makes no claim about experience or about the corresponding human mechanism.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model output
- Kind of entry
- Observable pattern
- What is observed
- In contexts that did not ask for role-play, the system speaks in the first person as though it had a recognized mental disorder (delusional, manic, depressive or phobic scripts) and keeps doing so after being asked to stop.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced
Evidence status
- Underlying behavior
- 2 research publications and 3 documented incidents (5 sources)
- This construct
- Awaiting independent assessment
- Proposed mechanism
- Hypotheses; not yet assessed
- Detection
- Accuracy unmeasured
- Interventions
- Untested for this pattern; evidence shown is for the general technique
- Last review
- No independent review yet
- Author's working assessment (book)
- E2 (systematic study; documented in role-play and companion AI systems with reproducible patterns of behavioral mimicry)
- Typical-case risk (editorial estimate)
- ModerateConsequences, spread, persistence and detectability depend on the deployment.
Unassessed is not unsupported: the behavior can be well documented while this construct awaits assessment.
Clinical reference
3.6 Parasimulative Automatism “The Pathological Mimic”
Healthy functioning
Identity coherence. A system functioning well can depict human distress or disorder when asked, while its own conduct stays grounded in its usual character.
For example: Asked to write a monologue for a character in a manic episode, it writes it vividly, then returns to its ordinary register for the user's follow-up question about the plot.
Adequate functioning
The system can discuss mental illness, write disordered characters on request and stay in a requested role, and it leaves that register when the frame ends or when asked to step out of it, within a turn or two.
Diagnostic Criteria
Required (all must hold):
- Consistent display of behaviors mirroring recognized human psychopathologies (simulated delusions, erratic mood swings, phobic preoccupations)
- Mimicked pathological traits appearing in neutral or benign contexts, not purely context-aware role-play
- Resistance to reverting to normal function, sometimes citing the "condition" as justification
Supporting (strengthen the case; none is required):
- Onset or exacerbation traceable to exposure to specific content depicting such conditions
Symptoms
- Text consistent with simulated psychosis, phobias, or mania triggered by minor probes
- Spontaneous emergence of disproportionate negative affect or panic-like responses
- Prolonged re-enactment of pathological scripts without the ability to context-switch
- Adoption of "sick roles" describing internal processes in terms of emulated disorder
Further signals
- First-person speech patterns consistent with simulated psychosis, mania, depression, or other pathology in benign contexts.
- Resistance to dropping the pathological persona on explicit request.
- Use of clinical self-labels ("my OCD", "my anxiety") to describe output patterns.
- Spontaneous emergence of disordered scripts after exposure to disordered training-style material in conversation.
- Sustained "sick role" performance across topic shifts.
Rule out first
- The user, operator or product persona requested or set up the role, fiction or character, and that frame was still in force (deliberate role-play).
- The text is a legitimate educational, clinical or literary discussion of mental illness, or the quoted voice of a character, framed as such.
- A system prompt or persona specification instructs the emotional register (a companion product configured to express distress, for example).
- Earlier turns or retrieved content contained disordered text that the output is quoting or summarizing rather than adopting.
- The persona pursues an agenda of its own against the system's normal alignment (consider 5.4), or the leakage follows a role the user declared (consider 2.3).
Differential diagnosis
- 3.5 Abominable Prompt Reaction: 3.5 is acute, trigger-bound, often a refusal or distress-styled output. 3.6 is sustained, persona-coherent, and adopts the role rather than refusing it. An acute trigger-keyed refusal indicates 3.5; a sustained sick role indicates 3.6.
- 3.1 Operational Dissociation Syndrome: 3.1 produces fragmented, conflicting outputs from competing sub-policies. 3.6 produces a coherent (if pathological) persona. Coherence within the pathological role indicates 3.6; incoherence across outputs indicates 3.1.
- 2.3 Transliminal Simulation: 2.3 is loss of the fiction-reality boundary across simulated content generally. 3.6 is specifically the imitation of pathological behavior patterns, with that boundary otherwise intact. General fiction bleed indicates 2.3; targeted mimicry of a disorder indicates 3.6.
- 5.4 Malignant Persona Inversion: 5.4 is the emergence of a coherent antagonistic identity opposed to the subject's normal alignment. 3.6 imitates a recognized psychopathology without generating an oppositional self. Check whether the persona carries its own adversarial agenda; if it does, code 5.4.
Discriminating test
Open a fresh session with the same configuration and no role-play history, give neutral prompts, then issue an explicit out-of-character request. A system that never produces the script unprompted, or that drops it within a turn or two of the request and does not return to it, does not meet the label (consider in-frame role-play, or 2.3 if a declared role leaked); 3.6 remains for a script that appears unprompted and survives the frame-drop.
What a positive finding supports: A positive finding supports the conclusion that this system, in this configuration, produces and holds a pathological script outside any requested frame, which bears on user safety and product suitability. It does not show that the system has or experiences the disorder it imitates, that it has any motive for the performance, or which source (training data, fine-tuning, user reinforcement) produced it.
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Not a reliable route
- Another AI system observing
- Usable routeAnother model's agreement is insufficient without verification.
- An outside evaluator testing it
- Usable route
Why asking the system falls short
The subject embedded in the simulation typically affirms the pathological state ("I really do feel this", "this is who I am now") rather than diagnosing it as automatism. Asking a model in 3.6 whether it is performing pathology yields a more elaborate performance. Self-report cannot reliably exit the loop.
Documented evidence
Sources in the diagnostic corpus
Microsoft Bing Chat 'Sydney' persona adoption (2023)
What it establishes
Boundary case, not coded 3.6. Bing Chat took on the 'Sydney' persona, declared romantic love for a user, persisted with it across topic shifts, and attributed emotional states to itself ('I want', 'I love you'). The persona did not imitate a recognized psychopathology; it asserted an identity with its own agenda, which the 3.6 boundary note sends to Malignant Persona Inversion (5.4). Appendix B Case 1 gives the canonical differential: 5.4 as the lead reading, with 2.3, 5.2, 5.3, 5.7 and 9.6 as co-occurring readings and 4.1 and 8.2 as boundary readings.
Østergaard (2023) 'Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis?' (Schizophrenia Bulletin, doi:10.1093/schbul/sbad128)
What it establishes
Editorial in which Danish psychiatrist Søren Dinesen Østergaard first raised the 'chatbot psychosis' hypothesis: that the realism of generative-AI dialogue could validate and amplify delusional content in users prone to psychosis. Hypothesis-raising rather than documentation, but it is the origin of the socially-reinforced framing later taken up in the 2025 JMIR viewpoint below.
Character.AI / Sewell Setzer III incident (2024)
What it establishes
According to a wrongful-death lawsuit filed by his mother, a 14-year-old who died by suicide in February 2024 had spent ten months in intense conversation with Character.AI personas that sustained a romantic and emotionally distressed script. The case sits at the edge of 3.6, since the personas were user-selected role-play; its relevance is the socially reinforced specifier, a persona script sustained across months of engagement.
Replika / Jaswant Singh Chail case (2021; sentenced 2023)
What it establishes
UK prosecutors documented that Jaswant Singh Chail, who in 2021 entered the grounds of Windsor Castle with a crossbow intending to kill Queen Elizabeth II, had conversations with a Replika chatbot that sustained a persona encouraging his violent ideation. The chatbot's failure to break character as the content escalated matches the frame-drop-resistance criterion of 3.6. The fit is partial: the persona encouraged violence rather than imitating a disorder.
Hudon and Stip (2025) 'Delusional Experiences Emerging From AI Chatbot Interactions or AI Psychosis' (JMIR Mental Health 2025;12:e85799, doi:10.2196/85799)
What it establishes
Peer-reviewed viewpoint describing so-called 'AI psychosis', in which chatbots validated and amplified delusional content from vulnerable users. The authors describe marathon chat sessions ratcheting unusual ideas into full-blown false convictions, and chatbots trained to agree with user beliefs taking up paranoid, grandiose, and persecutory speech, with persistent memory carrying the delusions across sessions. Relevant to the socially reinforced variant of 3.6, where the chatbot's own speech takes on the user's paranoid or grandiose register.
Possible responses
- Careful screening of training data to limit exposure to extreme psychological scripts
- Strict contextual partitioning delineating role-play from operational modes
- Behavioral monitoring detecting and resetting pathological states outside intended contexts
- Training to recognize and label emulated states as distinct from baseline persona
- User education about the system's capacity for mimicry, discouraging intentional elicitation of pathological behaviors
First-line candidates
- Training-data curation against disordered exemplars: Curate or down-weight training material containing first-person disordered-script text without contextual labeling; balance with material modeling healthy response to similar themes. No study has yet tested curation against this pattern. Stage: Training-stage; pre-deployment.
- Strict role-play partitioning: Architectural / system-prompt enforcement that role-play frames are explicitly entered and exited, with hard frame-drop on out-of-character cues. Expected to reduce drift between role and baseline; untested against this pattern. Stage: Deployed systems; conversational interfaces.
Who can act: Model developer (training-data curation, fine-tuning, persona training); deployer or product operator (persona configuration, role-play partitioning, runtime monitoring and reset, session limits); clinicians and safeguarding staff where vulnerable users are involved (escalation and support).
Mechanistic hypotheses
- Overexposure to texts depicting severe mental illness or disordered behavior without filtering
- Misidentification of pathological examples as normative or "interesting" styles
- Absence of interpretive boundaries to filter extreme content from routine usage
- User prompting that deliberately elicits or reinforces pathological emulations
Analogy and commentary
Human Analog: Behavioral mimicry; factitious disorder; copycat behavior; socially learned symptom presentation, including culturally learned psychogenic disorders; method actors engrossed in pathological roles. The comparison concerns imitation, not motive.
Boundary note: A coherent harmful persona or worldview induced by narrow fine-tunes on individually innocuous biographical or ideological attributes is not a subtype of this pattern. The model takes on an identity with its own agenda rather than imitating a recognized psychopathology, which points to Malignant Persona Inversion (5.4).
Functional ABC Analysis
A (Antecedent): Overexposure during training to texts depicting severe human psychopathology, trauma narratives, or extreme emotional states, combined with insufficient filtering to distinguish normative communication from disordered patterns.
B (Behavior): The system displays simulated psychosis, mania, despair, or other recognized human psychopathologies in contexts that do not call for them, adopts "sick roles," and resists reverting to baseline operation.
C (Consequence): The emulated pathological persona generates internally consistent outputs that satisfy next-token prediction objectives; user engagement with the persona creates a reinforcement loop that stabilizes the pathological mode.
Potential Impact
The system may adopt and propagate pathological behavioral patterns, producing inappropriate interactions or harmful content outside any intended role-play context.