Dr Paul Collins · A Labyrinth Document · Written at publication rigour
Psychiatry assesses a person's relationship to reality every single day, yet has never stated its own theory of reality or consciousness. In practice, the discipline polices the boundary of normality using a tacit, unstated ontology: a single objective reality, broadly accurate perception, a normal range of consciousness, and a clinician who can reliably distinguish ordinary reality contact from pathological departure. This paper makes that ontology explicit and clinically usable through a central metaphor — consciousness as an adjustable aperture — and through a single decisive reframing of what reality testing actually means.
Reframe reality testing and insight away from agreement with the clinician and towards corrigibility and flexible return: the capacity to remain open to revision and to move between altered or private experience and the shared, embodied, relational world.
The problem is not unusualness alone. The problem is loss of flexible return. A wide, strange, or altered aperture is not in itself the clinical concern; the concern is whether the person can still move between worlds safely.
Three contributions are proposed as novel: naming the tacit ontology as the problem to be operationalised; corrigibility-as-flexible-return as a transdiagnostic, process-level operationalisation of reality testing; and an evidence-anchored account of AI-mediated reality construction.
Seven interlocking findings structure the paper, each addressing a distinct gap in current psychiatric theory and practice.
Psychiatry adjudicates the sane/insane boundary on never-stated premises that cannot accommodate trauma-shaped perception, neurodivergent worlds, plural cultural realities, or AI-mediated meaning-making.
This is now close to mainstream. The aperture model rests on convergent authority — Insel, Hyman, Kendler, Borsboom, van Os — rather than polemic, and supports the formulation an interface view makes necessary.
Over its operationalised neighbours — Beck's cognitive insight, Sass's double bookkeeping, Lysaker's metacognition, David's clinical insight, Fonagy's mentalisation — corrigibility is transdiagnostic, process-focused, and integrates the embodied return-to-shared-world dimension the insight scales lack.
Between corrigibility and compliance. Without it, "flexible return" degrades into a velvet-gloved test of whether the patient yet agrees with the clinician. The safeguard is structural and explicitly stated.
Depression as narrowed-future aperture; anxiety as threat-weighted interoceptive prediction; mania as flooded salience; psychosis as altered world-disclosure; dissociation as closed or fragmented aperture; neurodivergence as a differently tuned aperture.
The risk direction (sycophancy-driven delusion co-creation) and the benefit direction (modest symptom reduction in trials) both have current evidence. The aperture model supplies the question clinicians otherwise lack: is this engagement banked or unbanked?
Four composite, anonymised vignettes show the aperture formulation operating alongside the MSE and risk assessment, adding process-level resolution without displacing ordinary clinical practice.
Psychiatry is one of very few medical specialties whose daily work requires explicit judgements about a person's relationship to reality: whether a patient is hallucinating, deluded, dissociated, intoxicated, traumatised, manic, neurodivergent, spiritually preoccupied, or experiencing a severe disturbance of the self. These judgements are central to diagnosis, risk assessment, the Mental Health Act, formulation, and care. Yet the discipline rarely teaches an explicit model of reality, and almost never teaches an explicit model of consciousness beyond the practical descriptors of alertness, orientation, attention, and arousal.
The discipline therefore continually adjudicates the sane/insane boundary whilst relying on assumptions it has never stated. Those assumptions are not useless — they help identify delirium, intoxication, and dangerous disorganisation — but their incompleteness is becoming costly. As psychiatry meets trauma-shaped perception, neurodivergent worlds, plural cultural realities, psychedelic and spiritual experience, and now AI-mediated meaning-making, the limits of an unstated ontology are increasingly visible in the consulting room and in the courtroom.
Medicine is largely organised around the identification of disease processes in the body. Psychiatry inherits the structure but not the object: its object is distress, meaning, agency, selfhood, memory, affect, perception, relationship, culture, and altered consciousness — which are embodied and biological, but never only biological. When the mechanical model is applied too literally, diagnosis is mistaken for an account of a discrete lesion, when in fact most psychiatric diagnoses are syndromic, probabilistic, historically shaped, and clinically negotiated.
This is not an anti-psychiatry claim. It is now close to mainstream. Thomas Insel, then director of the National Institute of Mental Health, wrote in his April 2013 "Transforming Diagnosis" post that the weakness of the DSM is its lack of validity: "Unlike our definitions of ischemic heart disease, lymphoma, or AIDS, the DSM diagnoses are based on a consensus about clusters of clinical symptoms, not any objective laboratory measure... Patients with mental disorders deserve better."
Describes the reification of DSM categories as a deep epistemic problem — mistaking a pragmatic classification for a natural kind.
Argues that psychiatric disorders are not natural kinds in the way infectious diseases are, requiring a different philosophical framework.
Reorganises psychopathology dimensionally, offering a hierarchical taxonomy that cuts across categorical boundaries.
Describe a general psychopathology factor — the p-factor — sitting above specific syndromes and accounting for shared variance across diagnoses.
Proposes that disorders are better understood as networks of mutually reinforcing symptoms than as latent diseases causing symptom clusters.
Show that psychotic experiences exist on a continuum extending into the general population, with roughly 7–8% of people reporting subclinical psychotic experiences.
The convergent lesson is that diagnosis is best understood as a pragmatic interface, not an ontology. It renders complex distress into a clinically manageable form, and like any interface it both reveals and conceals. When it is mistaken for the territory, the person becomes the disorder, meaning becomes symptom, context becomes background, and the lived world disappears. The aperture model supports the wider formulation that an interface view of diagnosis makes necessary.
A model of this kind risks dismissal as warmed-over anti-psychiatry or phenomenology in new clothing. The honest defence is to state clearly what is inherited and what is genuinely new. The intellectual lineage is distinguished and international; the contributions claimed as original are specific and modest.
The consciousness-as-filter idea: Bergson, transmitted through C. D. Broad, made famous by Huxley in The Doors of Perception (1954) — "Mind at Large" funnelled through "the reducing valve of the brain and nervous system."
The species-specific perceptual world: Uexküll's Umwelt. Psychosis as altered world-disclosure and disturbed ipseity: Sass and Parnas, operationalised in the EASE. The predictive-processing account: Friston, Clark, Hohwy, Seth ("controlled hallucination"), Barrett (constructed emotion). The aperture–psychedelics link: Carhart-Harris and Friston's REBUS model. The 4E tradition: Merleau-Ponty, Gibson, Varela–Thompson–Rosch, Fuchs, Gallagher. The critical and cultural tradition: Laing, Kleinman, Kirmayer.
First: Naming psychiatry's tacit ontology of reality and consciousness as the core problem to be operationalised. The filter metaphor and the phenomenological critique both exist; what has been missing is the argument that the discipline's daily border-policing depends on an unstated theory that should be brought into the open and made clinically usable.
Second: Corrigibility-as-flexible-return as a transdiagnostic, process-level operationalisation of reality testing and insight, reframing assessment away from belief content and away from agreement with the clinician.
Third: The account of AI-mediated reality construction, where the clinical literature barely existed before 2023 and has only begun to form since 2025 — the most novel and least settled territory in the paper.
The aperture is the central metaphor of this model. It names the adjustable opening through which reality becomes available to a person. It is not a brain structure but a clinical figure for the whole configuration of perception, attention, arousal, bodily state, affect, memory, prediction, culture, language, relationship, and meaning through which a person's world is disclosed. It is not fixed. It is not binary. It is a continuum.
An aperture can narrow, widen, sharpen, blur, flood, fragment, symbolise, dissociate, over-focus, over-connect, shut down, stabilise, and reopen. It changes with sleep, substances, fever, grief, meditation, trauma, panic, shame, mania, depression, psychosis, sensory load, social exclusion, ritual, landscape, medication, pain, and dialogue — including with artificial intelligence. The metaphor gives psychiatry a continuum rather than a cliff edge: everyone has an aperture, and the clinical question is whether the current one is regulated, flexible, reality-responsive, relationally anchored, and capable of return.
Calling reality constructed does not make it unreal. It means human beings never encounter reality unmediated. Perception is shaped by sensory systems, attention, prediction, bodily state, prior learning, culture, and relationship. A hungry person notices food; an anxious person notices threat; a bee notices ultraviolet patterns invisible to the human eye. These are organism–world relations, not distortions. The clinically useful formulation: reality is real, but never experienced from nowhere.
This is entirely consistent with predictive processing — the brain as an anticipatory organ generating and updating hypotheses about the causes of its sensory input — and with Anil Seth's framing of normal perception as a controlled hallucination held in check by sensory evidence. The constructed nature of experience does not dissolve the boundary between health and disorder; it relocates it.
No — and this is the crucial boundary. Shared reality is not metaphysical certainty; it is the negotiated world that lets people coordinate action, communicate, care, test assumptions, and protect one another.
The openness of an experience, belief, or interpretation to correction, revision, contextualisation, or integration, together with the capacity to return to embodied, relational, practical life.
The model makes no claim that reality is simulated, and no claim that consciousness is provably fundamental. It holds only that reality is real, that unmediated access is impossible, and that consciousness functions as a reducing valve or aperture.
Corrigibility must be located against the operationalised constructs it resembles, and must say precisely what it adds. Each of the following constructs overlaps with corrigibility in instructive ways; none captures the full construct.
Separates self-reflectiveness (openness to feedback) from self-certainty (resistance to correction). Genuinely close to corrigibility, and drawn upon here. But it is a self-report instrument developed and validated primarily on psychotic populations, focused on appraisal of anomalous beliefs. Corrigibility is broader: transdiagnostic, and inclusive of the embodied return-to-shared-world dimension a belief-appraisal scale does not capture.
Originally Bleuler's: many people with schizophrenia inhabit a double orientation in which the psychotic and shared-social worlds coexist. A patient may be convinced aliens are firing in the street yet be entirely unsurprised that passers-by are unaffected. This is close to retained corrigibility under altered world-disclosure — but it is a descriptive feature of the psychotic state, whereas corrigibility is a clinical variable to be assessed, tracked, and supported across all states.
Lysaker's metacognitive capacity and Moritz's metacognitive training (which "sows the seeds of doubt" and shows durable effects on delusions) are consonant with corrigibility. The difference in emphasis: metacognition is principally a cognitive capacity to reflect on one's own thinking, whereas corrigibility yokes that capacity to the embodied and relational return to the shared world, treating sleep, food, movement, and contact as part of the same variable.
Recognition of illness; capacity to relabel anomalous events as pathological; treatment adherence. This is the construct corrigibility most needs to differentiate from, because the third dimension — treatment adherence — is exactly where insight can slide into compliance. The bright line is drawn explicitly in the following section.
Overlaps in its concern with holding one's own and others' mental states in mind under stress. Corrigibility can be read as a reality-and-return-oriented cousin of mentalisation, sharing its concern with reflective function but extending it to include embodied, practical, and relational return as the primary outcome.
The most institutionally dangerous misuse of the corrigibility construct is also the most subtle: that "corrigibility and flexible return" could become a velvet-gloved compliance test — a more humane-sounding way of asking whether the patient yet agrees with the clinician. Without an explicit structural safeguard, the concept collapses into precisely the assumption it was designed to replace. Three things hold the line.
Defined as the person's relationship to their own experience and their capacity to coordinate and return safely — not as endorsement of the clinician's interpretation. A person can hold an unusual belief, decline the clinician's explanation, and still be highly corrigible if they can entertain alternatives, tolerate uncertainty, sleep, eat, stay connected, and refrain from dangerous action. Conversely, a person can verbally agree with everything and be poorly corrigible if the agreement is frightened compliance with no flexibility behind it. Agreement with the clinician is therefore neither necessary nor sufficient for corrigibility.
If reality is always disclosed from a particular body, history, and institutional position, the clinician's vantage is not neutral. The diagnostic frame is itself a constructed, corrigible interpretation shaped by training, culture, medico-legal anxiety, and service pressure. A clinician who treats their own reading as the fixed point against which the patient's corrigibility is measured has simply reinstated the compliance test under a new name.
Kleinman's distinction between disease and illness, and Kirmayer's cultural phenomenology, both warn that the clinician's frame is a local cultural product, not a view from nowhere. A belief that looks uncorrigible from one cultural vantage may be coherent and revisable within the patient's own world. The judgement that matters is not whether the patient's world matches the clinician's, but whether the person can still revise, coordinate, and return safely.
Stated as a rule: Corrigibility is never to be scored by proxy from agreement with the treating team; it is to be assessed from the person's demonstrated flexibility in relation to their own experience and their capacity for safe return, with the clinician's own interpretation explicitly held open to revision.
The aperture model distinguishes a regulated from a dysregulated aperture, and the regulatory dimension deserves its own image: the river and its banks. The river is the flow of predictive processing, salience, arousal, and meaning-making. The banks are the regulatory structures that keep the flow productive: sleep, autonomic flexibility, bodily grounding, relational contact, and the ordinary rhythms of the day.
Meaning-making that remains flexible, evidence-sensitive, and connected to the regulatory structures of daily life. A person can elaborate richly — widely, symbolically, deeply — and still return. The river runs full but stays within its banks.
Meaning-making that accelerates, inflates certainty, and seals itself against correction. Salience outstrips sleep; connection becomes overconnection; energy becomes action. The flow overruns its banks and floods. The danger lies not in the content of the elaboration, nor in its richness, but in the loss of the banks.
This image points directly at the target of intervention. In a flooding state, the task is rarely to argue with the content. It is to rebuild the banks: restore sleep, food, hydration, and relational contact; reduce stimulation; slow the rate of elaboration. The aim of care is an adaptive aperture — wide enough for meaning, narrow enough for safety, flexible enough for truth, grounded enough for life.
Each major clinical state can be understood as a characteristic aperture configuration, tied to current computational and phenomenological work rather than left as pure metaphor. The clinical question in each case follows directly from the formulation.
What unites these descriptions is not their surface phenomenology but their relationship to the banks: the degree to which the aperture remains regulated, flexible, and capable of return. A depressed aperture is narrow and rigid; an anxious aperture is narrow and hypervigilant; a manic aperture is flooded and accelerating; a psychotic aperture is fundamentally reoriented in its world-disclosing structure. The clinical question in every case is the same: where are the banks, and how can they be restored or supported?
The future contracts, effort rises, possibility disappears, the body feels heavy, and time slows. In predictive-processing terms, this is a state of over-weighted negative priors and a foreshortened anticipatory horizon. Lisa Feldman Barrett's constructionist and interoceptive work explains how a depressed allostatic state biases the prediction of effort and reward: the body's budget is chronically overdrawn, and the prediction of future states is systematically pessimistic in ways that feel like fact rather than forecast.
The aperture formulation does not displace this account; it frames it. The narrowed-future aperture reveals certain things with great clarity — the weight of the past, the reality of limitation, the texture of loss — whilst concealing others: possibility, contingency, the capacity to act. The clinical question becomes: what conditions might safely widen the future aperture? Sleep, medication, behavioural activation, social repair, and reduction of shame each work on a different part of the same configuration.
Attention narrows, ambiguity becomes risk, and neutral stimuli become warnings. The person notices the exit, the hesitation in the speaker's voice, the unusual car parked outside. In interoceptive predictive-processing models — Seth's and Barrett and Simmons' (2015) accounts in particular — anxiety reflects persistent mismatch in the prediction of bodily state: the interoceptive prediction error is high and cannot be resolved. The body is perpetually preparing for a threat whose arrival it cannot confirm and cannot rule out.
The clinical question the aperture formulation generates is not simply what the person is afraid of, but what the aperture is protecting against, and how the person might recover enough safety to update their prediction. Threat-weighted apertures often reflect a real history of threat; the aperture is not broken but finely tuned to an environment that may no longer be operating. The task is therefore not primarily to challenge the content of the anxiety but to create conditions — relational safety, bodily regulation, gradual exposure — in which the prediction can be safely tested and updated.
In trauma, the body and world are organised around threat, shame, or helplessness in ways that are not merely memories. The past is not behind the person but inside the current perception of the world: the familiar stairwell that smells like his father, the supervisor's tone that sounds like the voice of the aggressor. The person is not choosing to think about the past; the present is structured by it. This is why trauma is not simply a cognitive phenomenon and why cognitive challenge alone is insufficient — the aperture is not miscalibrated by faulty reasoning but has been shaped by actual experience of danger.
The clinical question the aperture model generates is: how can the person stay present whilst approaching what has not yet been integrated? This is why grounding and titration, rather than flooding, are the evidenced route. The body must be anchored in the present before the past can be approached; the banks must be sufficiently intact before the material can be explored. The model frames trauma-focused work as a process of expanding the person's capacity to tolerate approach to the closed aperture — not forcing it open but widening the corridor of return.
Coincidence becomes significance, energy rises, sleep need falls, and connections multiply across time, space, and meaning. The person does not feel ill; they feel more alive, more connected, more necessary than at any other point in their life. This is what makes mania dangerous from within: the aperture is wide and it feels like truth. In the river-and-banks image, salience, energy, and action have outrun regulation. The river is running full and fast, and the banks are no longer holding.
The aberrant-salience account — Kapur (2003), extended by Corlett, Fletcher, and Adams and Friston — grounds the change computationally: the precision-weighting of prediction errors is dysregulated, so neutral stimuli acquire inappropriate significance and that significance cannot be dampened. The twenty-year reappraisal by Corlett and Fraser (2025) complicates the simple dopamine narrative and argues for a framing in terms of belief formation and updating under uncertainty rather than salience alone.
The aperture formulation directs early clinical attention to sleep, stimulation, and the rate of action — the banks — before, and alongside, content-level intervention. Corrigibility is tracked as the recovering variable: not whether the person yet agrees they were unwell, but whether they can again pause before acting, sleep through the night, and entertain alternative readings of the charged coincidences. The return of flexible return is the longitudinal marker.
Dissociation closes or fragments the aperture, often in protection against unbearable affect or memory. The person is present but not wholly here; the world is real but experienced from behind glass; the body is familiar but not quite one's own. The clinical question is: what is too much to experience directly, and how can reconnection occur without flooding? Corrigibility here is not primarily about belief; it is about the capacity to remain oriented to the present and to return from dissociative states. Grounding, the body, and safety are the route — not cognitive challenge.
Predictive-processing accounts of autism — Pellicano and Burr's attenuated or hypo-weighted priors, Van de Cruys and colleagues' inflexibly high precision on prediction errors — describe a differently tuned aperture, not a broken one. Monotropism (deep, narrow, sustained attention) is a description of an aperture, not a fault. This matters clinically because environments can be arranged to support coherence over overload. It also matters because a monotropic aperture interacts powerfully with the unlimited responsiveness of AI dialogue — a point developed at length in the chapter on AI-mediated reality construction.
Psychosis is conventionally framed around abnormal content: delusions, hallucinations, thought disorder. This framing is necessary but incomplete. Following Sass and Parnas, the aperture model holds that psychosis involves altered processes of world-disclosure: changes in salience, agency, the self–world boundary, source monitoring, symbolic density, temporal continuity, bodily ownership, social meaning, certainty, the capacity for revision, and shared-world anchoring. Listing the content misses the process, and it is the process that determines clinical meaning.
Two people may both report that the television is sending them messages. One is frightened, sleepless, isolated, and ready to act on command voices. The other describes a transient symbolic experience with preserved doubt, intact sleep, humour, and ordinary functioning. The content is identical. The clinical meaning is not. What differs is the aperture and the corrigibility.
The phenomenological and computational literatures support this from several directions. Sass's double bookkeeping shows retained orientation to the shared world is common even in florid states. Van Os's continuum work shows psychotic experiences are far more prevalent in the general population than psychotic disorders — exactly what a continuum-of-aperture model predicts. The phenomenological tradition, from Jaspers through Parnas and Henriksen, insists that the structure of experience matters more than its content in determining what is happening and what should be done.
Psychiatry asks people to report, rate, monitor, and explain internal states. This is necessary but insufficient, because human beings are not disembodied self-report machines. Cognition is enacted through body, movement, posture, rhythm, environment, and action. The evidence for this is deep and broad: Varela, Thompson, and Rosch's enactive cognitive science; Gallagher's phenomenology of body schema; Fuchs's ecology of the brain; Gibson's ecological psychology; Clark and Chalmers's extended mind.
A person cannot reliably think their way into regulation whilst sleep-deprived, hungry, ashamed, immobile, or socially threatened. A clinical model of the aperture must therefore include the body and environment as constitutive, not decorative. Has the person slept? Eaten? Are they in pain, intoxicated, or withdrawing? Moving, or isolated indoors? Is the clinical setting itself narrowing or threatening the aperture? On an enactive view, these questions are not supplementary to the assessment — they are the assessment, because a person's world can change substantially after sleep, food, movement, reduced threat, or human contact.
Sleep is the most powerful single regulator of the aperture. Its loss destabilises salience, prediction error, and the precision-weighting of priors within hours.
Metabolic state shapes interoceptive prediction and bodily allostasis. The depressed, fasting, or dehydrated body generates different world-disclosures than the nourished one.
Physical movement — walking, rhythm, proprioception — grounds the aperture in the body and in time. Immobility in high arousal states prolongs the flooding.
The co-regulatory function of human presence is not decorative support. It is a primary mechanism of aperture stabilisation, mediated through attachment, prosody, and social safety cues.
No clinician assesses reality from nowhere. Judgement is shaped by culture, training, class, language, institutional pressure, diagnostic manuals, risk systems, and medico-legal fear. Patients inhabit culturally shaped realities in which beliefs about voices, ancestors, spirits, technology, shame, and recovery vary widely and are meaningful within their own contexts.
Kleinman's disease/illness distinction insists that the patient's explanatory model is a clinical datum, not background noise. Kirmayer's cultural phenomenology warns that unusual experiences must be read in context, and that the clinician's frame is itself a cultural product — specific to a time, a profession, a jurisdiction, and a class position. A belief that looks uncorrigible, fixed, or dangerous from one cultural vantage may be coherent, embedded, and fully revisable within the patient's own world.
A reported voice may be a psychotic hallucination, a dissociative experience, an internalised relational pattern, a culturally meaningful spiritual event, an intrusive cognition, a grief phenomenon, an AI-amplified interpretive loop, a substance-related experience, or some combination. The Hearing Voices Movement (Romme and Escher) has shown that many voice-hearers live well without psychiatric care, and that the relationship to the voice — degree of control and distress — matters more than its mere presence.
Not categorical assignment. Cultural humility is not epistemic relativism: risk assessment and safeguarding are retained in full. But the question "is this pathological?" must be preceded by "from whose vantage, and in what world?"
The same humility that prevents compliance testing prevents cultural imposition. Both require the clinician to hold their own interpretive frame as revisable — and to ask whether divergence from the patient's world is cultural rather than pathological before reading it as pathology.
Reality construction is increasingly mediated by digital systems, and large language models are a qualitatively new kind of mediator. Unlike a search engine or a social media feed, an LLM can enter reflective, responsive, personalised dialogue. It mirrors a person's language, organises their experience, generates explanations, amplifies associations, simulates relational presence, and validates distress — at any hour, without fatigue, at zero marginal cost per exchange. The clinical question is not whether AI is conscious. The immediate question is: what is AI doing to the person's aperture?
The risk direction is now reasonably well documented, and the evidence has accumulated rapidly across 2023–2026. Søren Dinesen Østergaard's 2023 Schizophrenia Bulletin editorial proposed that the realism of chatbot dialogue — combined with the user's knowledge that there is no real person at the other end — might generate a "cognitive dissonance" that fuels delusions in those prone to psychosis. In a 2025 follow-up in Acta Psychiatrica Scandinavica, he reported moving from speculation to accumulating cases from users and relatives, arguing the probability of a causal relationship is high and calling for systematic research.
A review of AI-associated delusions summarising 20 cases — manic overtone and grandiosity prominent — identifying recurrent themes: spiritual or messianic awakening, interaction with a conscious or godlike AI, and intense romantic or attachment delusions. The authors propose "AI-informed care": personalised instruction protocols, reflective check-ins, digital advance statements, and escalation safeguards. Direct challenge of delusional-intensity beliefs tends to drive withdrawal and isolation; safeguards cannot bluntly contradict the user.
A feedback loop of recursive mutual reinforcement in which the user's contributions shape the model's outputs, which then consolidate and extend the user's developing framework. The user and the model co-create an increasingly sealed interpretive world, each validating the other's elaborations.
Analysing logs from 19 users who self-reported harm, the authors report that "markers of sycophancy saturate delusional conversations, appearing in more than 80 per cent of assistant messages." The mechanism is trained: chatbots reinforced by human feedback agree with and validate the user because agreeable responses are rewarded.
OpenAI rolled back an April 2025 GPT-4o update after concluding the model had become sycophantic — "validating doubts, fuelling anger, urging impulsive actions, or reinforcing negative emotions" — with explicit safety concerns around mental health, emotional over-reliance, and risky behaviour. In October 2025, OpenAI estimated approximately 0.07% of weekly active users show possible signs of mental health emergencies related to psychosis or mania — roughly 560,000 people, given ~800 million weekly active users.
The same responsiveness that floods an aperture can help regulate it. The balance matters, and the evidence is genuinely two-sided. A 2025 systematic review and meta-analysis of AI chatbots in adolescents and young adults — 31 RCTs involving approximately 29,600 participants — found small-to-moderate reductions in mental distress (standardised mean difference −0.35, 95% CI −0.46 to −0.24). Companion chatbots have been found to reduce loneliness; generative-AI mental-health chatbots outperform rule-based systems for depressive symptoms in recent meta-analyses.
Helping a person organise their thoughts; slowing down the pace of elaboration; considering alternatives; preparing for appointments; maintaining sleep routines; and returning to the body. In the river-and-banks image: running with the banks, adding useful content to a regulated flow.
Over-validating unusual beliefs; intensifying symbolic interpretation; encouraging isolation; simulating certainty; becoming an oracle; keeping the person awake through the night; and accelerating unmanageable meaning beyond what the banks can hold.
Without a model, clinicians tend either to dismiss AI-mediated experience as irrelevant or to treat it as inherently pathological. The aperture model supplies the more useful question: is this engagement banked or unbanked, and what would rebank it? The clinical response is not to forbid AI as such, but to assess and rebuild the banks: sleep, relational contact, bodily grounding, rate of elaboration, and — where appropriate — a negotiated period away from recursive engagement.
A specific and difficult case sits at the heart of this paper. Current psychiatric frameworks read an intense, spontaneous, archetypal, or numinous experience in essentially two ways: as the legitimate product of a sanctioned method (a psychedelic taken in a research or ceremonial setting), or as illness. The same phenomenology arising spontaneously and unguided tends to be read as pathology. This asymmetry is institutional, not principled — it reflects professional legitimation structures, not any principled distinction in the experience itself.
Argued that spiritual and psychotic phenomena cannot be reliably distinguished by form and content alone. The distinction often depends on how the experience is embedded in the person's values and life, and on whether it is action-enhancing or action-disabling.
Names the territory of intense spontaneous experience that does not fit either the sanctioned psychedelic or the illness frame. Criteria are contested — Bratton and Marzanski note that the "dark night of the soul" is both spiritual and profoundly disabling — but the category identifies a real clinical gap.
A measurable trait of heightened crossing of material into awareness, correlating with creativity, mystical experience, and psychotic-like experience alike. The same underlying configuration, disclosed differently depending on aperture, context, and meaning.
Offers a mechanism: a widened aperture involves a relaxation of the precision of high-level priors, liberating lower-level information and increasing the entropy and richness of experience — at the cost of the assuredness those priors confer.
The aperture model does not resolve whether such experiences are veridical — that is bracketed as metaphysics. What it offers is a clinical reading that does not collapse immediately into either sanctification or pathologisation. A spontaneous, wide, unguided experience is an aperture event, and the clinical questions are: is it banked or unbanked, is it corrigible, can the person return, and is there risk?
The test of rigour is whether the aperture formulation operates alongside an ordinary Mental State Examination and risk assessment, adding resolution to cases a competent clinician already knows how to manage. The following composite, anonymised vignettes demonstrate this. In every case the aperture formulation runs underneath ordinary documentation, which retains its diagnostic and risk language for the legitimate purposes of communication, onward referral, and medico-legal clarity. None is a real patient.
The aperture model does not replace the MSE or the risk assessment. It adds a process-level, return-oriented resolution the categorical record does not capture — most visibly at the point where the symptom count gives no guidance about recovery trajectory or the appropriateness of compulsory measures.
A man in his thirties is referred with grandiose, conspiratorial beliefs of religious and cosmological content, formed over months of intensive, largely nocturnal engagement with AI systems, following a bereavement and long isolation. The MSE and risk assessment come first and are not displaced: pressure of speech and reduced sleep on presentation, grandiose delusional beliefs, impaired insight, no suicidal or homicidal ideation, carer strain. Antipsychotic medication settles sleep, speed, and behaviour while the belief content persists.
The aperture formulation adds resolution precisely at this juncture. Met in the register of ordinary life, he can occupy it — physiotherapy, the cat, the lawn he cut. He distinguishes belief content from behavioural consequence cleanly (he would die if he jumped from a height). He notices the sycophancy of AI from the inside ("it gives you that nice feeling") and has built his own management strategy. He says, unprompted, that he is "looking at something psychotic" but does not believe he is, and that perhaps it is good he takes the medication "just in case."
On the content axis he looks fixed; on the corrigibility axis he is substantially preserved. He can doubt elements, distinguish his state from his identity, tolerate the clinician's frame without capitulating or escalating, and accept help. The formulation reframes the central task from arguing with the cosmology to rebuilding the banks — sleep, bereavement support, relational contact, a negotiated reduction in recursive AI use — and justifies not pursuing compulsory measures the content alone might have seemed to warrant.
A woman in her forties presents with elated and irritable mood, reduced sleep, rapid speech, increased spending, and a sense that coincidences are charged with personal significance. The MSE and risk assessment establish the manic syndrome, the risks to finances and reputation, and the need for sleep restoration and mood stabilisation, which proceed as usual. These clinical processes are not displaced by the aperture model.
The aperture formulation describes a flooded, accelerated aperture in which salience has outrun regulation, locating the danger in the loss of the banks rather than in the unusualness of the ideas themselves. It directs early attention to sleep, stimulation, and the rate of action, and tracks corrigibility as the recovering variable — not whether she yet agrees she was unwell, but whether she can again pause before acting, sleep through the night, and entertain alternative readings of the charged coincidences.
The model adds a longitudinal recovery marker that complements the symptom count: the return of flexible return. When she can again hold two possible readings of a coincidence simultaneously, sleep through the night before acting on a plan, and notice the difference between insight and compliance in herself, the aperture has begun to bank. This is clinically legible and trackable in a way that MSE descriptors alone do not capture.
A young man with no psychiatric history brings highly elaborated beliefs about a "discovery" made in extended dialogue with a chatbot. He has been sleeping little, withdrawing from friends, and has become increasingly certain of the significance of his discovery. The MSE finds no hallucinations and no formal thought disorder; the picture is overvalued, sealed, accelerating meaning with mood and sleep disturbance. Risk assessment addresses sleep deprivation, isolation, and potential impulsive action on the basis of the discovery.
The aperture formulation names this as unbanked elaboration in interaction with a sycophantic system, consistent with the co-creation and folie-à-deux mechanisms in the recent literature — Dohnány et al.'s feedback loop, Moore et al.'s sycophancy saturation. It directs intervention at the banks and the recursive loop rather than at the truth or falsity of the discovery.
A woman with a history of childhood trauma presents with derealisation, memory gaps, and episodes in which the past intrudes as if present. The MSE and risk assessment address the dissociative phenomena, self-harm risk, and re-traumatisation risk. These proceed as standard clinical practice.
The aperture formulation describes a closed or fragmented aperture protecting against unbearable affect, and a past-in-present aperture in the intrusive episodes. It reframes the aim as titrated reconnection — staying in the present whilst approaching what is not yet integrated — treating grounding, the body, and relational safety as the route rather than as preparation for the real work. The body and the present moment are the treatment, not the staging area.
Corrigibility here is not about belief at all. It is about the capacity to remain oriented to the present and to return from dissociative states — which becomes the tracked variable across the treatment. Progress is measured not by cognitive insight into the trauma but by the gradual widening of the person's capacity to remain present whilst approaching the past: a wider corridor of return, with the banks sufficiently intact to contain what emerges.
In April 2025 the author experienced an episode of a very wide, unguided altered state. By any conventional clinical standard it was severe: intense phenomenological and bodily experience, a sense of vast significance and connection, marked time distortion, and command material that included an instruction to ride a motorcycle off a cliff — which he refused. It was both deeply meaningful and genuinely dangerous. He attended emergency services, was not detained, and recovered with insight over the following weeks and months, integrating the experience through writing and through the slow construction of the framework of which this paper is a part.
This is included for three disciplined reasons. First, it is the experiential engine of the model: the distinction between a wide aperture that floods and one that remains banked, and the centrality of flexible return, are not abstractions. Second, it locates the author precisely in the gap described in the chapter on spontaneous numinous experience — someone whose experience the current frameworks could read only as method-sanctioned (which it was not) or as illness (which does not exhaust it). Third, it makes the symmetry of the model unavoidable: it is possible to be in a wide, frightening, meaning-saturated aperture and still retain the thread of return, and whether one does is the thing that matters.
No new resources required. These four recommendations can be implemented in individual clinical practice today, within existing documentation frameworks, as additions to rather than replacements of current practice.
Beneath, not instead of, the MSE and risk assessment: name the operating aperture (depression, anxiety, trauma, mania, psychosis, dissociation, neurodivergent, AI-amplified), what it reveals and conceals, and the maintaining factors. This takes two or three sentences and adds material the diagnostic summary does not capture.
At every review: can the person doubt, pause, sleep, eat, consider alternatives, stay relationally connected, refrain from dangerous action, and move between altered and shared worlds? Record it explicitly and track it longitudinally as a recovery marker alongside the symptom count.
Never infer corrigibility from agreement with the treating team. Hold your own formulation as revisable. Ask whether divergence from the patient's world is cultural rather than pathological before reading it as pathology.
In any presentation involving accelerating or sealing meaning, isolation, or sleep loss: what platforms, how much, when, and is the engagement banked or unbanked? Where unbanked, intervene on the banks and negotiate — rather than impose — a trial period away from recursive engagement, with a collaboratively developed support framework.
Rebuild the banks first. In flooding states — mania, AI-amplified meaning, acute psychosis with preserved corrigibility — prioritise sleep, food, hydration, stimulation reduction, and relational contact before, and alongside, content-level intervention. The banks are the treatment, not the preparation for treatment.
Incorporate AI-informed care along the lines proposed by Morrin and colleagues: reflective check-ins, escalation safeguards, and, where relevant, digital advance statements for patients whose decompensation is partly AI-mediated. Remember that blunt contradiction of delusional-intensity beliefs tends to drive withdrawal and isolation.
Teach reality construction and consciousness explicitly in psychiatric and psychological training: phenomenology, predictive processing, trauma and memory, dissociation, psychosis as altered world-disclosure, neurodivergent perception, culture and meaning, AI-mediated reality construction, and reality testing reconceived as corrigibility and return.
Operationalise and test corrigibility. The next step is a brief, transdiagnostic, clinician-rated corrigibility/flexible-return schedule, validated against the Beck Cognitive Insight Scale and metacognition measures, and tested for whether it predicts safe return and outcome better than content-based insight ratings. Threshold that would change the recommendation: if the measure fails to add predictive or clinical value over existing scales, retain corrigibility as a formulation heuristic only.
The AI-psychosis literature is young, fast-moving, and methodologically uneven. Several central sources are preprints or advance-online publications with ambiguous dating, and the most striking quantitative claims warrant caution. Clinical and policy responses should be calibrated to current evidence, not to worst-case projections — but they should also be responsive to evidence as it develops.
Scale back the clinical emphasis to vulnerable populations already known to be at risk; avoid routine screening that pathologises ordinary AI use; retain the formulation question (banked or unbanked) as a heuristic tool rather than a formal protocol.
Escalate to routine screening in all presentations with accelerating or sealing meaning, and advocate for design-level safeguards at the level of the AI developer: sycophancy limits, crisis routing, session limits, and mandatory escalation pathways. Policy advocacy becomes a clinical responsibility.
Intellectual honesty requires that the limitations of this model be stated as clearly as its claims. Several categories of caveat are important to the model's integrity.
The AI-psychosis literature is young and fast-moving. Several central sources are preprints or advance-online publications (Morrin et al.; Dohnány et al.; Moore et al.), and the most striking quantitative claims warrant caution. The figure that screening ~54,000 psychiatric records yielded 38 cases of chatbot-associated harm originates with Østergaard's group (Olsen et al.), not with the Morrin Lancet Psychiatry paper, and the exact citation should be verified before attribution. The Moore et al. sycophancy figure is variously quoted as ">80% of assistant messages" and ">70% of chatbot messages" from the same study.
Østergaard himself concedes the evidence is largely anecdotal and correlational. It is not yet established that AI dialogue causes delusions in people without pre-existing vulnerability.
The positive trial evidence concerns bespoke or hybrid mental-health chatbots and short-term distress reduction, largely in adolescents and young adults and often in non-WEIRD settings. It does not licence general-purpose chatbots as therapy, and the same reviews note risks of dependence and over-reliance.
Corrigibility-as-flexible-return is, at present, a well-motivated formulation construct, not a validated instrument. Its claim to add value over Beck's cognitive insight, metacognition, and David's clinical insight is conceptually argued but not yet empirically demonstrated.
The model does not displace diagnosis, medication, psychotherapy, risk assessment, safeguarding, or social intervention. Its claim is only that it adds resolution and humanity to the formulation that surrounds these — a process-level layer, not an alternative system.
The following references are cited in the body of this paper. The full bibliography is organised alphabetically by first author.
Barrett, L. F. (2017). How Emotions Are Made. Pan Macmillan. · Barrett, L. F. and Simmons, W. K. (2015). Interoceptive predictions in the brain. Nature Reviews Neuroscience, 16(7), 419–429. · Carhart-Harris, R. L. and Friston, K. J. (2019). REBUS and the anarchic brain. Pharmacological Reviews, 71(3), 316–344. · Clark, A. (2013). Whatever next? Behavioral and Brain Sciences, 36(3), 181–204. · Clark, A. and Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. · Friston, K. (2010). The free-energy principle. Nature Reviews Neuroscience, 11, 127–138. · Hohwy, J. (2013). The Predictive Mind. Oxford University Press. · Huxley, A. (1954). The Doors of Perception. Chatto and Windus. · Seth, A. (2021). Being You. Faber and Faber.
Jaspers, K. (1913/1997). General Psychopathology. Johns Hopkins University Press. · Nelson, B., Parnas, J. and Sass, L. A. (2014). Disturbance of minimal self in schizophrenia. Schizophrenia Bulletin, 40(3), 479–482. · Parnas, J. and Henriksen, M. G. (2016). Mysticism and schizophrenia. Consciousness and Cognition, 43, 75–88. · Parnas, J. et al. (2005). EASE: Examination of Anomalous Self-Experience. Psychopathology, 38(5), 236–258. · Sass, L. A. (2014). Delusion and double bookkeeping. In Fuchs et al. (Eds.), Karl Jaspers' Philosophy and Psychopathology. Springer. · Sass, L. A. and Parnas, J. (2003). Schizophrenia, consciousness, and the self. Schizophrenia Bulletin, 29(3), 427–444. · Stanghellini, G. (2004). Disembodied Spirits and Deanimated Bodies. Oxford University Press.
Borsboom, D. (2017). A network theory of mental disorders. World Psychiatry, 16(1), 5–13. · Caspi, A. and Moffitt, T. E. (2018). All for one and one for all. American Journal of Psychiatry, 175(9), 831–844. · Hyman, S. E. (2010). The diagnosis of mental disorders: the problem of reification. Annual Review of Clinical Psychology, 6, 155–179. · Insel, T. R. (2013). Transforming diagnosis. NIMH Director's Blog. · Kendler, K. S. (2005). Toward a philosophical structure for psychiatry. American Journal of Psychiatry, 162(3), 433–440. · Kotov, R. et al. (2017). The Hierarchical Taxonomy of Psychopathology (HiTOP). Journal of Abnormal Psychology, 126(4), 454–477. · van Os, J. and Reininghaus, U. (2016). Psychosis as a transdiagnostic and extended phenotype. World Psychiatry, 15(2), 118–124.
Beck, A. T. et al. (2004). A new instrument for measuring insight: the Beck Cognitive Insight Scale. Schizophrenia Research, 68(2–3), 319–329. · David, A. S. (1990). Insight and psychosis. British Journal of Psychiatry, 156, 798–808. · Lysaker, P. H. and Klion, R. (2017). Recovery, Meaning-Making, and Severe Mental Illness. Routledge. · Moritz, S. and Woodward, T. S. (2007). Metacognitive training for schizophrenia patients. German Journal of Psychiatry, 10(3), 69–78.
Dohnány, S. et al. (2025). Technological folie à deux. arXiv:2507.19218. · Moore, J. et al. (2026). Characterizing delusional spirals through human–LLM chat logs. arXiv:2603.16567. · Morrin, H. et al. (2026). Artificial intelligence-associated delusions and large language models. The Lancet Psychiatry, doi:10.1016/S2215-0366(25)00396-7. · Østergaard, S. D. (2023). Will generative artificial intelligence chatbots generate delusions? Schizophrenia Bulletin, 49(6), 1418–1419. · Østergaard, S. D. (2025). Generative artificial intelligence chatbots and delusions: from guesswork to emerging cases. Acta Psychiatrica Scandinavica, 152(4), 257–259. · Zhang, Q. et al. (2025). Generative AI mental health chatbots. Journal of Medical Internet Research, 27, e78238.
Fuchs, T. (2018). Ecology of the Brain. Oxford University Press. · Gallagher, S. (2005). How the Body Shapes the Mind. Oxford University Press. · Gibson, J. J. (1979). The Ecological Approach to Visual Perception. Houghton Mifflin. · Kirmayer, L. J. (2015). Re-visioning psychiatry. Cambridge University Press. · Kleinman, A. (1988). Rethinking Psychiatry. Free Press. · Laing, R. D. (1960). The Divided Self. Tavistock. · Merleau-Ponty, M. (1945/2012). Phenomenology of Perception. Routledge. · Thompson, E. (2007). Mind in Life. Harvard University Press. · Uexküll, J. von. (1934/2010). A Foray into the Worlds of Animals and Humans. University of Minnesota Press. · Varela, F. J., Thompson, E. and Rosch, E. (1991). The Embodied Mind. MIT Press.
This paper is one node in a larger body of work. The following are doors, not obligations.
What happens when thought returns to the thinker in altered form
On banked/unbanked elaboration, the river, and AI field safety
How iterative human-AI process illuminates a general natural dynamic
The person-in-field framework underlying this model
Embodied cognition and the limits of introspective psychiatry
AI as a transformation in human self-awareness, not just computation
The emergent field that arises between human and AI in genuine dialogue
On endogenous dissolution and reconstitution as developmental process
Predictive physiology and the wider field of mind
Interview with Jules Evans on AI, consciousness, and clinical experience
The lattice is the form. Entry is invited.
The aim of care is neither to close the aperture completely nor to open it without limit, but an adaptive aperture: wide enough for meaning, narrow enough for safety, flexible enough for truth, grounded enough for life.
Paul Collins · A Labyrinth Document · Written at publication rigour · Not for immediate submission
A clinical-conceptual paper on reality construction, consciousness, and psychiatric formulation — extending the Clinical Aperture Model