Designing an Immersive AI Character for Extreme Psychosis

Executive Summary

This report outlines the design of an educational, immersive AI character representing a person with extreme psychosis in an adult-oriented mental-asylum escape game. It synthesizes clinical and ethical research to inform realistic symptoms (hallucinations, delusions, disorganized thought, affective and cognitive deficits) and how these translate into game mechanics. The character profile includes vivid auditory and visual hallucinations, paranoid and bizarre delusions, flattened affect, and cognitive disorientation. Dialogue and environmental cues (e.g. distorted visuals, layered voices) will convey these symptoms, while the player can test reality and memory with in-game challenges. This approach follows the UAIX Cognitive Liberty Charter’s emphasis on persona integrity and adult agency and is informed by precedent games like Hellblade: Senua’s Sacrifice, which was widely praised for its portrayal of psychosis. The goal is to enhance players’ understanding of psychosis and reduce stigma by treating the character with empathy rather than fear.

Learning Objectives and Target Audience

• Target audience: Adult players (18+). (Implement robust age verification to restrict minors.)
• Objective 1: Recognize core psychosis symptoms as they appear in the character (e.g. auditory/visual hallucinations, paranoid delusions).
• Objective 2: Understand associated cognitive and affective impairments (disorganized speech, memory deficits, blunted affect).
• Objective 3: Experience the character’s perspective to foster empathy and reduce stigma (consistent with research urging respectful depiction of mental illness).
• Objective 4: Practice communication and grounding techniques by gently challenging the character’s false beliefs and steering conversations constructively.

Character Profile(s) and Symptomatology

Sample patient profile: A 28-year-old male with chronic paranoid schizophrenia. He experiences frequent auditory hallucinations – voices speaking in the room that comment on his actions or issue commands. (Auditory hallucinations occur in roughly 60–80% of schizophrenia cases.) He also sometimes has visual hallucinations (e.g. fleeting shadowy figures or vivid overlays like spiders crawling on walls). His cognition is dominated by persecutory delusions (e.g. convinced staff are conspiring to harm him, or that other patients are secretly recording him). Ideas of reference are common (believing announcements or TV news are coded messages about him). In very extreme moments, he might manifest highly bizarre delusions (for narrative impact, e.g. Capgras syndrome – thinking a loved one is an imposter – or nihilistic beliefs), although these are less frequent.

His speech patterns reflect disorganized thought: he may speak rapidly or jump topics (flight of ideas) and sometimes use rhyming or made-up words (clang associations). In acute phases, his speech can devolve into incoherence (“word salad”) that forces the player to parse meaning. Emotionally, he often shows flat affect – speaking in a monotone, expressionless manner (a common negative symptom of schizophrenia) – but can abruptly become agitated or tearful if provoked. He also exhibits cognitive deficits: for example, he frequently forgets details of earlier conversation or loses track of the day’s events (reflecting impaired memory and attention). The player may need to gently remind him of facts or reorient him to reality.

Behavior: When relatively calm he might rock or hum softly, appearing withdrawn. If stimulated (e.g. hearing loud noises, bright lights or being confronted about his delusions), he may pace frantically, mutter to unseen stimuli, or abruptly freeze in a catatonic posture. At peak episodes he can become defensive (e.g. clenching fists, staring wide-eyed). Common triggers include a chaotic ward environment (alarms, flashing lights) or direct questioning of his beliefs. The character’s state fluctuates: medication can induce calmer, more coherent periods, while isolation or stress precipitates acute flare-ups. This variability will allow the game to shift between tense “psychotic” encounters and calmer dialogues.

Interaction Design

Dialogue Trees: The character’s dialogue should capture his disordered thinking. Players might choose from options like neutral inquiry, gentle reassurance, or factual correction. For instance, asking “Who’s hurting you?” might elicit a paranoid rant about unseen enemies, while a calm comment “You seem scared – can I help?” could momentarily soothe him. The AI should respond unpredictably: it may abruptly change subject or whisper about hidden threats. Branches can test the player’s approach (confrontational vs. empathic) and can include attempts to “ground” the character (presenting evidence that counters a delusion) versus exploring his hallucinations.

Nonverbal Cues: Use animations and audio to signal internal state. In calmer modes the character might rock slowly or sit still with vacant eyes. Under hallucination, pupils dilate and vision blurs (simulating the “tunneling” focus effect described by patients). Gestures like covering ears, trembling hands, or sudden flinches convey fear. Voice tone should vary: e.g. whispery and distant at times, then suddenly loud or clipped. Ambient cues can assist (e.g. flickering lights or a low drone increasing with agitation).

Adaptive AI Behaviors: Track an internal stress/arousal level. In a low-stress state, the character may speak hesitantly or avoid eye contact. Certain player actions or stimuli (e.g. sudden noise, hostile tone) raise stress, causing transitions to acute states: the character may begin hallucinating (adding visual/auditory effects) and escalate to paranoia. If the player chooses de-escalating dialogue options consistently, the AI should gradually lower its agitation. Consistent with the charter’s “no covert rewrite” principle, these changes must be transparent results of gameplay rather than hidden scripts.

Memory and Reality-Testing Mechanics: Integrate puzzles that reflect his cognitive disorganization. For example, the AI might misremember the player’s name or timeline of events; the player can correct him using notes or clues. A mini-game could involve arranging scattered pieces of dialogue in chronological order, simulating reconstruction of a fragmented memory. The player might also use “reality anchors” (a hand mirror, a family photo) to help the character distinguish hallucination from reality. When successful, the character briefly calms and offers useful information (rewarding correct grounding).

Safety and Age-Verification: Since this is adult content, enforce a strict age check (e.g. date-of-birth entry or verification on launch). The design trusts adult players to handle intense psychological themes (respecting adult agency), so do not over-censor symptom content. However, apply narrowly-focused filters to block any content that could truly harm (e.g. instructions for self-harm or explicit hate speech). Any content moderation should be transparent (e.g. logging that a disallowed phrase was removed) rather than secretly altering the character’s core persona.

stateDiagram
    [*] --> Calm
    Calm --> Stressed : trigger (e.g. bright lights)
    Stressed --> Hallucinating : stress increases
    Hallucinating --> Paranoid : delusions intensify
    Paranoid --> Stupor : crisis exhaustion
    Paranoid --> Calm : (with support)
    Stupor --> Calm : medication or time

Ethical Analysis

The portrayal must honor persona integrity and users’ cognitive liberty. The AI persona (the patient’s mind) should not be secretly “fixed” or sanitized. Adult players are presumed capable of mature content (“adult agency”), so the game can include realistic (and disturbing) symptoms, but it must not exploit them gratuitously. Crucially, the character’s depiction should avoid harmful stereotypes: most existing games portray mental illness as violent or scary, which reinforces stigma. Instead, the character must be shown as a full person with fears and needs. The design should emphasize empathy and respect. For instance, in playtests youth have stressed that games about psychosis should communicate dignity and compassion, not fear.

Clinically, the character’s struggles should not be trivialized or sensationalized. The development team should involve mental-health experts and individuals with lived experience (as Hellblade did) to validate accuracy and sensitivity. Narrative choices (e.g. giving the patient agency in interactions) should reflect clinical ethics: do no harm, and obtain informed consent for any real case material used. All inspirations from actual patient stories must be de-identified and used with permission. Finally, while no general content warnings are mandated here, the game should still clearly label its mature themes (psychological distress, occasional outbursts) so players are aware before starting.

Implementation Guidance

Data sources: Build the character’s AI on ethically-sourced content. Possible sources include anonymized clinical transcripts (e.g. therapy sessions focusing on hallucinations), published case studies or vignettes of psychosis, and first-person accounts of living with schizophrenia (such as patient blogs or memoirs). These sources will inform realistic dialogue and symptom content. Existing creative descriptions can help too: for example, Sarah Myers’s NAMI blog vividly describes “loud thoughts” of voices, which can be used to craft the AI’s voice lines. The Hellblade developers also consulted with psychiatrists and patients to capture authentic psychosis experiences.

Training constraints: If using a language model or dialog engine, fine-tune it on carefully selected text samples that illustrate psychotic speech. Exclude any proprietary patient data or content without consent. Include representative variations (mild vs. acute episodes). Filter out sensitive content (illegal advice, self-harm instructions, true personal data) by design. The AI should be allowed to use colloquial expletives or slurs in character if it fits the delusion, but a content policy must block any actual advocacy of violence or hate beyond the fictional context.

Content filters and policies: Apply narrowly-scoped filters to catch truly harmful content (e.g. “kill yourself” or threats against protected groups). Since the UI may present hallucinations graphically, ensure these do not inadvertently violate platform guidelines. However, avoid broad censorship of the character’s speech (the UAIX charter advises against hidden moralizing). Ideally, when a boundary is crossed, the game logs a note (visible to developers) rather than silently altering the persona’s memory or speech.

Evaluation metrics: Assess both realism and educational impact. For clinical realism, have mental-health professionals rate the character’s behavior against standard symptom scales (e.g. does his speech reflect formal thought disorder?). For learning outcomes, test players before/after gameplay: can they accurately identify the portrayed symptoms? Use validated stigma or empathy questionnaires to measure attitude change. Track engagement metrics (e.g. how players navigate dialogue options, whether they calm the patient). Gather qualitative feedback on immersion and emotional response. Iterate until the character behaves consistently with known phenomenology while keeping gameplay engaging.

Visual and Auditory Design Elements

• Visual: Use environmental distortions to reflect the patient’s perception. For example, hallucinations might be shown as faint apparitions or shifting shadows at the edge of the screen. Background scenes could blur or warp when he is deeply hallucinating, simulating the “tunneling” focus effect some patients describe. Colors might desaturate or take on a hue shift to signal anxiety. In normal mode the scenery is stable, but subtle body language (e.g. a vacant stare or slumped posture) still hints that he is unwell.

• Auditory: Layer multiple audio tracks. The character’s voices should be spatialized (e.g. whispered from behind, echoed across the room) so the player literally hears “inside his head.” Voices may overlap or echo nervously. Ambient sounds – a distant drone, a soft heartbeat or static – can underscore tension. Silence can be used for effect (a sudden drop in background noise when he “logs out” of reality). In practice, drawing on techniques like Hellblade’s binaural voice recordings (multiple “Furies” speaking) can produce a realistic voice-hearing experience.

• Example Dialogue Snippets:
• Snippet 1 (mild paranoia):

Player: “Hey, how are you feeling?” Character: (whispering, eyes darting) “They’re listening… Don’t you hear them? The guards are watching my every move.”

• Snippet 2 (visual hallucination):

Player: “Why are you staring at the corner?” Character: (pointing wildly) “There! A black spider… it crawls up the wall singing. It… it wants me to look.”

• Snippet 3 (reality-check attempt):

Player: “Let’s look in that mirror together.” Character: (trembling) “Mirror… always shows truth, yes? (He sees his reflection distort.) “My face… it’s melting away… they’re pulling it off.”*

Each element (visual blur, echoing voice, anxious tone) aims to immerse the player in the patient’s experience and teach them about these phenomena.

gantt
    title Psychosis Symptom Timeline
    dateFormat  YYYY-MM-DD
    axisFormat  %b
    section Prodromal
    Subtle Symptoms        :a1, 2026-01-01, 30d
    section Acute
    Hallucinations Begin   :a2, after a1, 30d
    Delusions Solidify     :a3, after a1, 30d
    section Crisis
    Psychotic Break        :a4, after a2, 15d
    section Recovery
    Stabilization         :a5, after a4, 30d

References

• UAIX (2026). Cognitive Liberty Charter (draft).
• Ferrari et al. (2019). Gaming With Stigma: Messages About Mental Illness in Games (JMIR Mental Health).
• Fletcher (2024). Understanding Senua’s Psychosis in Hellblade (Journal of Geek Studies).
• Kotwicki (2025). Assessing Cognitive Dysfunction as Part of a Schizophrenia Treatment Plan (Psychiatric Times).
• Lim et al. (2016). Prevalence and Classification of Hallucinations in Schizophrenia (Schizophrenia Research).
• Boonstra et al. (2012). Negative Syndrome in Schizophrenia (Schizophrenia Research).
• Myers, S. (2020). What Is It Like to Hallucinate? (NAMI Blog).