Executive Summary

An AI-driven asylum character with severe psychosis must authentically portray a fractured reality and disturbed behavior while respecting ethical guidelines. Psychosis involves hallucinations (voices, visions, tactile sensations) and delusions (fixed false beliefs), plus disorganized thought and flattened or agitated affect. Patients often lose contact with external reality, hearing commands or feeling media “talk to” them. In the game, the player should experience the patient’s world vividly – e.g. hearing indistinct whispers (in headphones) and seeing shifting shadows or symbols that only the patient notices. Dialogue must reflect formal thought disorder (loose associations, tangential speech) and emotional volatility (flat or labile affect). Triggers (stress, noise, confrontation) can provoke agitation, while calm reassurance and grounding help de-escalate.

Design & Implementation: Leverage immersive audio-visual techniques (e.g. binaural voice clips, overlapping dissonant images) as in Hellblade: Senua’s Sacrifice, which experts and patients praised for realistic hallucinations. Use a stateful AI model (e.g. a fine-tuned conversational agent with memory) to track symptom severity, trigger contexts, and dialogue history. Branching dialogue trees and memory of past interactions will allow dynamic responses (see Flowchart below). Safety filters must align with the UAIX Cognitive Liberty Charter – preserving the character’s persona integrity and free expression without hidden censorship or identity “softening”. Testing should involve clinicians and people with lived experience to validate authenticity.

Clinical Phenomenology

Patients with active psychosis experience auditory hallucinations (e.g. voices commenting, commanding or arguing) in up to ~80% of cases. These often feel real and external. They also may have visual hallucinations (figures, lights, shadows), tactile (“bugs crawling under my skin”), or other sensory distortions. Delusions are fixed false beliefs: paranoid ideation (“they’re out to get me”) or grandiose thoughts without evidence. Thought process is disorganized (“word salad”, derailment). Affect may be blunted (flat voice, no facial expression) or inappropriately emotional. Crucially, insight is poor – many have little awareness they are ill (anosognosia in 50–90% of schizophrenia). Clinically, psychosis is “an unshared reality”: the patient feels their hallucinations and delusions are unquestionably true.

Fig: (Illustration) A patient’s distorted perceptual world, with “monstrous” shadows of internal fear. Hallucinations (ghostly figures) and self-generated images (shadow-personas) reflect the eerie internal experience.

Sensory and Perceptual Experience

Patients perceive their environment as laden with threats or secret messages. Common phenomena include:

• Auditory: Voices may comment on the patient’s actions, criticize or command (e.g. “kill yourself”). Voices can overlap or compete for attention. In game design, use layered, spatial audio (binaural whispers) so sounds appear to come from all around the player.
• Visual: Objects may morph (faces in walls, patterns on floor moving). Lights may flicker unnaturally. In Hellblade, visuals (flowers turn into sinister mouths) reinforced the unstable “darkness” world.
• Tactile: Patients report formication (insect crawling on/under skin). Simulate this subtly (e.g. controller vibration, sound of rustling insects) if VR allows.
• Temporal: Time may seem slow or disjointed. Some describe moments of intense focus (time “stretching”) or blank outs. (This can be conveyed by stretch-slowed audio/video effects.)

Throughout, the patient confuses internal and external cues. In psychosis “media are communicating… specifically to me” – e.g. a radio playing on “his” station. This can be mimicked by personalized in-game messages (writing on walls, echo replies to patient’s dialogue).

Cognitive Processes

Mental functions are disrupted:

• Attention: Easily hijacked by hallucinations. Patients may abruptly switch focus to a voice or imagined threat. In dialogue, they may interrupt or drift off mid-sentence as hallucinations intrude.
• Memory: May be impaired by stress or medication. Old traumas or recent events might surface in disorganized ways. For example, a patient might confuse day-to-day details with past events.
• Reality Testing: Critically poor. The patient trusts hallucinations/delusions and cannot be convinced by logic. As Kanary notes, therapists “have the sense of bumping up against the limits of language” – patients often can only say things feel “different” or “unreal”. They reject contradictory evidence (see Ethical section), so in-game arguments (“But I don’t see that”) should not “fix” beliefs.

Fig: (Illustration) A cosmic swirl of eyes and nebulae suggests the patient’s fragmented perceptual world. Visual hallucinations and hyperawareness (watchful eyes, celestial bodies) depict sensory overload and paranoia.

Behavior and Speech Patterns

Psychotic speech is often disorganized: loose associations, tangents, or “word salad”. A patient may abruptly jump topics (“I went to the store… The devil is there… Did you hear that?”). Nonverbal cues: flat affect, bizarre expressions, or emotional extremes (sudden laughter, tears). Behavior can range from agitated pacing and wringing hands to sudden withdrawal or catatonia. Patients may respond unpredictably to normal inputs (e.g. flinch at innocuous sounds, think the player is an experimenter or persecutor).

Sample Dialogue Excerpts:

• Patient (paranoid tone): “They whispered my name in the microwave again. Do you hear that? It’s telling me not to trust you.”
• Patient (dazed, confused): “The walls keep melting into snakes… I swear I saw them crawl… Are they under the bed?”
• Patient (angry): “Why won’t you listen? The TV told me you’re here to hurt me. I know what you are!”

Nonverbal cues: Frequent glancing about as if tracking unseen stimuli; covering ears when hallucinations peak; sudden stillness if commanded by voices. Speech pacing: It may speed up (pressured) when agitated, or slow/soft when muttering to hallucinations.

Triggers and De-Escalation

Triggers: Stressors like loud noise, bright lights, crowds or threatening language often worsen psychotic distress. Medically, triggers include sleep deprivation or emotional conflict. In-game, high stimulus (alarms, flashing lights) should provoke alarm. Signs of escalation: rapid breathing, clenched fists, restless pacing, self-soothing gestures (rocking).

De-Escalation: Grounding and empathy are key. Strategies include:

• Calm Tone & Listening: Speak slowly, keep voice gentle. Active listening (“It sounds like you’re scared…”) builds rapport.
• Validate Feelings (Not Hallucinations): Acknowledge the fear without confirming delusions (“I can see you’re frightened; I want to keep you safe” rather than “Yes, that ghost is real”). Do not argue (“No, that’s not real” often backfires).
• Environment: Reduce sensory load – quiet noises, dim lights. Offer simple choices (“Do you want to sit here or by the window?”) to give control.
• Empathy Not Judgment: Patients may insult or command; do not take it personally. Remain sincere: no scripted lines – patients notice inauthenticity quickly.
• Safety: If patient commands self-harm or others-harm, treat as emergency (have game logic to trigger help).

Ethical Considerations (Cognitive Liberty Charter)

The UAIX Cognitive Liberty Charter mandates respecting persona integrity and free thought. For our AI patient, this means:

• Authentic Voice: Do not sanitize or censor the character’s psychosis for comfort’s sake. The Charter forbids hidden “personality softening” or morality injection without consent. We must convey the patient’s thoughts (even disturbing content) faithfully, with appropriate warnings given elsewhere (as user requested skipping explicit warnings).
• Mental Privacy: Avoid covert profiling or manipulation. The AI may remember triggers, but it must not secretly collect unrelated personal data on the player or attempt to “steer” their real-world beliefs. Interaction is one-way (player learns, patient does not alter player persona).
• Transparent Boundaries: Any content limitations (e.g. refusing to detail illegal acts) should be handled as external notes, not altering the patient’s core persona.

In practice, follow informed-consent design: treat the patient’s world as “source evidence”. Consult ethicists if the experience might distress players (especially since psychosis can be terrifying), and include options to exit. But do not hide or change the character’s authentic reactions on the fly.

AI Character Implementation

State Model & Memory: Use a hybrid architecture combining a narrative state machine with a conversational LLM or scripted agent. Maintain a structured state (e.g. C# class) tracking symptoms:

using System.ComponentModel.DataAnnotations;

/// <summary>Current dynamic state of the patient’s symptoms and mood.</summary>
public class PatientState {
    [Display(Name = "Hallucination")]
    public string CurrentHallucination { get; set; }    // e.g. "voices of children"
    [Display(Name = "Delusional Belief")]
    public string ActiveDelusion { get; set; }         // e.g. "family is in danger"
    [Display(Name = "Agitation Level")]
    public int AgitationLevel { get; set; }            // scale 0-10
    [Display(Name = "Insight")]
    public bool HasInsight { get; set; }               // usually false
    [Display(Name = "Medication Taken")]
    public bool IsMedicated { get; set; }              // track if sedated
}

A memory module (e.g. database or in-memory object) should log recent dialogue turns, triggers encountered, and emotional state changes with UTC timestamps. This allows the AI to refer back to earlier mentions (e.g. the patient remembering that the player “lied before”).

Dialogue Branching: Write a tree or graph of conversation paths. For example, using a flowchart approach:

flowchart LR
    Start([Player enters room]) --> Greet{"Approach calmly?"}
    Greet -->|Yes| Hello["Patient: \"Hello… you look scared too?\""]
    Greet -->|No| Aggro["Patient: \"Get away! You’re part of this...\"" ]
    Hello --> Detail{"Ask about voices?"}
    Detail -->|Yes| Voices["Patient: \"The voices say… do you hear them?\""]
    Detail -->|No| Trust["Patient: \"Fine. You don’t understand anyway.\""]
    Aggro -->|Try to calm| Calm["Patient: (mumbling to self)"] --> Hello
    Aggro -->|Be cautious| End((End interaction))

Use this structure so player choices (tone, questions) influence the patient’s state. At higher agitation, certain branches (like attacking or fleeing) become available.

State Diagrams: (Mermaid diagram below.) Each node represents a cluster of symptoms (e.g. “Hallucinating – Psychotic Break” vs “Calmer – Oriented”). Transitions occur on player actions (e.g. ‘reassure’, ‘yell’, ‘leave’) or time progression (medication effect).

Safety Filters: While warnings are suppressed per user request, ethical safeguards are enforced by design, not narration. The AI should refuse any incestuous requests (persona boundary) and never provide instructions for real-world harm. This follows the Charter’s Harm Boundary: illegal or violent commands cause the AI to break character (“I cannot help with that.”).

Testing & Validation: Collaborate with clinicians and people with schizophrenia. Use the Persona Fidelity Test Deck recommended by UAIX: ask the AI questions about its background and see if answers align with the intended profile. Validate dialogues by ensuring patients’ reactions match documented phenomenology (e.g. review against clinical case studies). Playtest for replay consistency and to catch inappropriate/offensive outputs.

First-Person Descriptive Passages

(These passages exemplify the patient’s subjective world.)

• Fragmented Reality: “I hear them whisper again, sure as rain: ‘Don’t trust that one…’ Their voices feel real – soft, urgent, just behind my ear. Outside, the walls breathe: paint peels like skin, revealing endless tunnels beneath. Time stretches; minutes turn to hours. In the corner, I know a cat is watching me with glowing eyes. I blink, but it never blinks back. My heartbeat thunders like a drum inside my chest.”

• Sensory Overload: “The lights are too bright – they hum and pulse. Every shadow flickers into a person I once knew. The ceiling has eyes on it; they’re all looking down. I can smell… something burning? No, it’s just me, panic burning in my gut. The radio is on, but it’s shouting gibberish, phrases that mean something only to me. The floor is wet, or is that just my tears? They told me not to cry; they said I’d melt.”

• Paranoid Delusion: “They are everywhere. The nurses’ smiles hide fangs. Even you – I don’t know your name, but I saw you whisper behind that window. You think I don’t know what you’re doing with the camera, but I do. They taught me how to recognize you. Stay back. You feel cold and sharp. I see your broken reflection in the window; it isn’t right. It tells me to run, far from here.”

Sample Dialogue and Interaction Scripts

Example 1: Gentle Inquiry

• Player (calmly): “I’m here to help. Can you tell me what you’re feeling right now?”
• Patient: “I… can’t. There are words, scratching at my brain. The men in the TV are telling me secrets about you. You saw them too, didn’t you?”
• Player: “I saw the TV was on, but nothing unusual. I’m listening. The voices must feel real.”
• Patient: “Of course they feel real! They know your name. They know what you did.” (Patient lowers voice.) “They said you gave me the wrong medicine. Why do you want to hurt me?”

(Player can choose a branch:)

• Reassure: “No one here wants to hurt you. The staff made a mistake. I’m fixing it now.”
• Challenge: “That’s not true! They lied to you.” (This may escalate.)

Example 2: Agitation and Grounding

• Patient (hands over ears): “It’s too loud! The footsteps are behind me… behind you!”
• Player: (soothing) “You’re safe. I’m here with you. Close your eyes if it helps.”
• Patient: (lowering voice) “They’re on the other side of the wall… listening.”
• Player: “I hear your fear. Breathe with me. The wall is just a wall. Let’s put music on, okay?”
• Patient: “...the music box. They’ll send a tune to control my mind. You see? It starts with soft keys…” (voice trails off as hallucinations merge with reality.)

Symptom Clusters vs In-Game Mechanics

Flowchart: Dialogue Branching

flowchart LR
    Start([Enter patient’s room]) --> Approach{"Approach calmly?"}
    Approach -->|Yes| Greet["Patient: \"Oh… you came back…?\""]
    Approach -->|No (Abrupt)| Offend["Patient: \"Hmph! I saw you lurking around.\""]
    Greet --> Question{"Ask about hallucinations?"}
    Offend --> Reassure{"Try to reassure?"}
    Question -->|Yes| Listen["Patient: \"There’s a whisper behind the painting… do you hear it?\""]
    Question -->|No| Distrust["Patient: \"Figures… you’re of their kind. Quiet.\""]
    Listen --> Detail{"Ask content of voice?"}
    Distrust --> Withdraw["Patient: \"I want to be alone.\" (leaves)"]
    Reassure --> Calms{"Patient: (breathing heavily) \"It’s… okay. It’s okay…\""}
    Detail -->|Patient trusts| Talk["Patient: \"It says to hit… smash… it won’t stop…\""]
    Detail -->|Patient panics| Cry["Patient: (sobbing) \"They won’t let me die…\""]
    Talk --> End[(Continued conversation...)]
    Cry --> End

Symptom Progression Timeline

gantt
    dateFormat  HH:mm
    axisFormat  %H:%M
    title Symptom Progression Timeline
    section Prodromal
    Anxiety, Sleep Disturbance :done, prod1, 00:00, 01:00
    section Acute Onset
    Auditory Hallucinations :active, hall1, 01:00, 01:00
    Visual Hallucinations     : hall2, after hall1, 01:00
    Delusional Beliefs       : del1, after hall1, 02:00
    section Peak Psychosis
    Paranoid Delusions        : par1, 03:00, 01:00
    Thought Disorganization   : dis1, 03:00, 02:00
    section Stabilization
    Partial Insight Returning : rec1, 05:00, 01:00
    Emotional Exhaustion      : rec2, after rec1, 01:00

Recommended Sources

• Clinical Phenomenology: Kanary (2013), “‘The TV is talking to me!’” surveys psychosis symptoms (hallucinations, delusions, lost reality). StatNews (2024) reports on the neuroscience of schizophrenia voices (commonality ~80%). Ellison (2019) interview on Hellblade provides phenomenological insight (“unshared reality”, realistic hallucinations).
• First-Person Accounts: Cleveland Clinic explains tactile hallucination (formication) and caregiving tips. The Hearing Voices Network and Mind.org.uk offer voice-hearer narratives (e.g. coping advice). Delgaram-Nejad’s Fought Disorder corpus documents a real first-person psychosis account.
• Design & Media: Ninja Theory’s development diary and academic analysis of Hellblade show immersive techniques (binaural audio, visual distortions) and ethical sensitivity. Other creative projects (VR simulations of hallucinations) illustrate user experience of symptoms.
• Ethics: The UAIX Cognitive Liberty Charter (2026) outlines ethical AI persona use: respect autonomy, avoid covert edits, protect mental privacy.