OLYA
designing trust in AI-mediated diagnosis
Overview
Reimagining mental health evaluations through VR pass-through & AI clarity.
Mental health assessment feels inaccessible—expensive, intimidating, and requiring you to admit you need help before you even know what's wrong.
Traditional clinical assessments demand vulnerability with strangers in unfamiliar settings. But introducing AI raises new questions: How do you design an AI system people will trust with their mental health? What does trustworthy AI actually look like?
So, what exactly did we make?
Our final design combines a pass-through VR evaluation with a clear two-layer report.
People move through the evaluation in their own physical space while tasks appear one at a time, and clinicians get a structured summary that supports their workflow.
Once the psychiatrists and providers have completed their review, patients receive a version of the same information written in language they can act on as well as the diagnosis.
Everything lives inside MyChart, so nothing about adoption relies on a brand-new tool.
Before we get ahead of ourselves, how did we get here?
We worked within the realities of clinical practice from day one. All tasks had to be based on real evaluations, nothing invented or speculative. AI couldn’t diagnose; it could only help organize and clarify information that was already there.
The experience needed to work reliably across different environments, and whatever we built had to integrate cleanly into existing healthcare tools. Given the time constraints, we had to design an interaction rhythm in VR that felt steady and approachable without overcomplicating it.
Our goals followed directly from those constraints: make the process easier to move through, communicate clearly with both audiences, keep the information model consistent, and build something a hospital could realistically* adopt.
*as realistic as this can get
Three critical design decisions.
Decision 1: Why AR over VR?
CONSTRAINT:
Diagnostic assessment takes 60-90 minutes. Users need to feel safe and comfortable at home.
THE TRADE-OFF:
VR offers deeper immersion but creates physical isolation from the environment. AR provides less immersion but maintains situational awareness.
OUR DECISION:
AR allows users to:
- See exits and their surroundings (safety)
- Integrate real furniture (comfort—sit on their couch)
- Respond to interruptions (pets, family) without full disconnection
- Feel less "trapped" in an assessment experience
LEARNING:
In high-stakes contexts, "good enough" presence with maintained
agency beats maximum immersion.
Decision 2: Translating Clinical Protocols to AI Delivery
CONSTRAINT:
The Wechsler Adult Intelligence Scale (WAIS-IV) requires
precise, standardized administration. Even slight variations
in how questions are asked can invalidate results.
THE CHALLENGE:
How do you maintain clinical standardization while making the
AI feel supportive rather than robotic?
OUR APPROACH:
We studied actual WAIS-IV administration protocols and designed
a three-layer system:
1. PROTOCOL LAYER (invisible to user)
• AI follows exact timing, question sequencing, scoring rules
• No deviation from clinical standards
2. GUIDANCE LAYER (visible when needed)
• "Take your time—there's no wrong answer"
• Contextual encouragement without influencing responses
3. ADAPTATION LAYER (smart transitions)
• "Let's take a break" after long sessions
• Adjusts pacing based on fatigue indicators
Decision 3: Making AI Diagnosis Legible
CONSTRAINT:
"Black box" diagnosis feels illegitimate. Users need to understand
how the AI reached its conclusions.
THE TRADE-OFF:
Full transparency (showing all test scores, algorithms) is
overwhelming. Too little transparency feels opaque.
OUR APPROACH: Progressive Disclosure
LEVEL 1 - HEADLINE (immediate clarity)
"You've been diagnosed with Attention-Deficit/Hyperactivity
Disorder (ADHD)"
LEVEL 2 - CRITERIA (show the reasoning)
"This diagnosis is based on DSM-5 criteria..."
[Expandable sections for each criterion met]
LEVEL 3 - DATA (for those who want details)
"View full WAIS-IV results"
[Link to complete score report]
LEVEL 4 - CONVERSATION (make it interactive)
"Talk to Olya about your results"
[AI assistant answers questions about the diagnosis]
Interaction
model
Three Phases of AI-Mediated Diagnosis
We designed Olya's interaction model around three distinct phases, each requiring different AI behaviors and trust mechanisms.
But the real design challenge isn't just making assessment cheaper or more accessible—it's about trust.
THE CORE QUESTION:
How do you design an AI system that people trust with consequential decisions about themselves, when they can't see the human expert behind it?
This became our design problem: creating a diagnostic experience that feels as legitimate as sitting across from a psychologist, while running entirely on consumer AR hardware and AI systems.
Phase 1: Assessment (AI as Neutral Administrator)
DESIGN PRINCIPLE: Calm authority
During testing, the AI needs to feel:
- Neutral (not influencing responses)
- Clear (users know exactly what to do)
- Professional (this is a real clinical test)
INTERACTION DETAILS:
- Voice guides through tasks with exact WAIS-IV phrasing
- Visual instructions appear in peripheral view
- Minimal personality—this isn't a chatbot, it's a clinician
- "Recording" indicator shows test is being captured
INNOVATION:
3D spatial puzzles that leverage AR in ways traditional tests
can't—rotating objects, exploded views, spatial reasoning tasks
that feel native to the medium.
Phase 2: Results (AI as Interpreter)
DESIGN PRINCIPLE: Clarity with empathy
After testing, the AI shifts from neutral to supportive. Users
are receiving consequential information about themselves.
INTERACTION DETAILS:
- Results appear in MyChart (familiar, trusted platform)
- Headline diagnosis first—no burying information
- "Learn More" expands to DSM-5 criteria
- "Emerging" language for potential secondary diagnoses
- Full WAIS-IV report downloadable as PDF
INNOVATION:
Conversational AI interface: "Talk to Olya about your results"
Users can ask questions like:
- "What does this mean for my daily life?"
- "Should I be worried about the OCD indication?"
- "How accurate is this diagnosis?"
The AI provides evidence-based responses while acknowledging
its limitations.
Phase 3: Next Steps (AI as Care Coordinator)
DESIGN PRINCIPLE: Integration over isolation
A diagnosis is only valuable if it leads to care. Rather than
leaving users stranded with results, Olya integrates directly
into the healthcare system.
INTERACTION DETAILS:
- "Schedule appointment" links to in-network providers
- Educational resources specific to diagnosis
- Family history questionnaire (AI identifies patterns)
- Medication information if prescribed
- Progress tracking for future sessions
INNOVATION:
Treating the diagnostic report as a living document rather than
a static result. Users can return to Olya to:
- Track symptom changes
- Prepare for doctor appointments
- Understand medication effects