A Thought Experiment: What if AI in Healthcare Behaved Like a Consultant?
Healthcare AI faces a paradox: clinical documentation keeps getting heavier despite "smart systems," and AI assistance often increases cognitive load rather than reducing it.
This thought experiment proposes an architectural inversion: What if AI systems in healthcare were structurally incapable of creating, modifying, or storing medical records?
Not as policy. As architecture.
Like a consultant who reads everything, reasons deeply, gives their assessment, and leaves no note.
Not a better writer. A better reader.
AI assisting chart comprehension without becoming part of the medical record.
In practice, almost nothing is ever removed from medical records.
So records accumulate:
- Twenty-page clinical notes
- Impossible chart reviews
- Documentation that is medically complete but cognitively impossible to navigate
For the first time, we have systems that can read and reason over these impossible records.
But current implementations try to:
- Listen to encounters
- Write clinical notes
- Update the EHR
- Become part of documentation flow
This creates three collision layers:
Records aren't just memory — they're evidence.
Question: "What did the AI remove that we might need later?"
Clinical judgment has always meant "what I see, remember, and write."
AI seeing patterns across thousands of records creates discomfort: "If it sees more than me, am I still deciding?"
Every previous "helpful" technology added burden.
EHRs were supposed to help. They became a source of burnout.
Institutional memory: "Tech makes this worse."
The Absolute Invariant:
The system must not be able to create, modify, store, or remember medical truth.
Not "shouldn't." Must not be able to.
If the design allows it, the design is wrong.
Not an "AI scribe."
A Session-Bound Clinical Reasoning Engine.
Characteristics:
- Stateless
- Non-persistent
- Non-exportable
- Non-citable
- Non-archivable
It exists only while the clinician is looking at the chart.
Like a calculator. Or a stethoscope.
When you close it, it's gone.
- EHR, clinical notes, audit trail
- Nothing here changes. Ever.
- Satisfies HIPAA, liability, compliance
What it does:
- Context compression on demand
- Temporal stitching (10 years of notes → "what actually happened")
- Treatment trajectory extraction
- Contradiction detection across years
- Pattern surfacing across visits, medications, symptoms
Critically: It stores nothing. Leaves no trace. Produces no new record.
Like a radiologist reading documentation instead of scans.
You look. You reason. You discard.
Output is never: "Here's the new note"
Output is: "Here's what you may want to look at before you write your note"
AI = pre-reading assistant, not writer
Explicit rules:
- AI outputs are non-archivable
- AI summaries cannot be stored in EHR
- AI reasoning cannot be cited in court
- AI is classified as temporary cognitive tool, not system of record
This removes 80% of legal fear instantly.
The AI does not remember patients.
It reads the record fresh every time.
Meaning:
- No cross-patient contamination
- No latent memory risk
- No PHI persistence
- No model "learning" from patients
Every reasoning session is stateless.
| Current Fear | This Design Removes It |
|---|---|
| "AI will write wrong things into record" | AI never writes |
| "What if AI deletes something important" | AI never deletes |
| "What if AI remembers PHI" | AI stores nothing |
| "What if AI replaces clinician judgment" | AI only pre-reads |
| "We don't trust tech after EHR burnout" | This reduces reading burden, doesn't add writing burden |
During a session, the clinician can ask:
- "What is the true timeline here?"
- "Summarize the treatment trajectory"
- "Show me contradictions across years"
- "Why was this medication changed three times?"
- "What matters in these 10 pages that I must not miss?"
The AI reads. Reasons. Shows.
The clinician closes the panel.
Nothing remains.
All current systems try to:
- Listen
- Write notes
- Update EHR
- Become part of documentation flow
This system:
- Never touches documentation flow
- Only touches cognitive flow
- Is ephemeral by architecture
- Cannot persist anything
That's a category healthcare hasn't named yet.
This thought experiment synthesizes principles from existing architectural frameworks:
- Session-bound state reasoning
- Adaptive control without persistent memory
- State-aware regulation layer
- Structurally non-representable memory
- Privacy through irreversibility
- Cannot reconstruct specific instances
- Trust through inspectable architecture
- Contestable by design
- Safety through constraint, not alignment
- Governance must precede persistence
- Structural properties over intent
- Containment architecture over alignment
This architecture does not provide:
- Automated documentation (intentionally)
- Clinical decision-making (preserved for humans)
- Diagnosis assistance (not the focus)
- Universal EHR integration (implementation detail)
It provides: Cognitive assistance without documentation burden or legal liability.
- Real-time EHR read access (no write)
- Secure session isolation
- Stateless processing architecture
- Auto-purge on session close
- Explicit non-archival policy
- Legal classification as "temporary tool"
- Clear clinician authority boundaries
- Audit trails for access, not content
- Opt-in per session
- No persistent profile
- Clinician-initiated queries only
- No autonomous suggestions
The Consult Model reflects:
- Familiar clinical metaphor (consultants read everything, then leave)
- Ephemeral expertise (present during decision, absent after)
- Non-ownership (consultant doesn't write in the chart)
- Authority preservation (attending makes final decisions)
Alternative framings:
- Clinical Cognitive Navigation System (CCNS)
- Ephemeral Clinical Reasoning Layer
- Session-Bound Chart Review Assistant
This is a thought experiment, not a product.
It exists to:
- Reframe how AI assistance in healthcare could work
- Question assumptions about AI "writing" in clinical settings
- Propose architectural constraints that address legal/authority/workflow fears
- Demonstrate how existing AI safety patterns apply to healthcare
It is not:
- A implementation specification
- A commercial proposal
- A replacement for existing clinical workflows
- A solution to all healthcare AI challenges
Healthcare is approaching an inflection point where:
- Documentation burden is unsustainable
- AI capability is sufficient for reasoning tasks
- Legal/compliance frameworks are unclear
- Trust in healthcare technology is low
This thought experiment asks:
"What if we stopped trying to make AI write better notes, and instead helped humans reason over notes no human can fully process anymore?"
Not replacing clinical judgment.
Supporting it — by carrying the cognitive weight that governance structures created but humans were never built to carry.
- ARCHITECTURE.md — Technical layer specifications
- INVARIANTS.md — The constraints that make this work
Collaborative Design:
- Leena Thomas (systems architect)
- Thea/ChatGPT (architectural synthesis)
- Claude (documentation and governance framing)
MIT License — This is a thought experiment, freely shareable and adaptable.
This thought experiment applies systems architecture principles to healthcare AI challenges.
For a complete catalog of related research:
📂 Research Index
Architecturally related:
- Connector OS — Session-bound state reasoning
- SMA-SIB — Structurally non-representable memory
- Doctrine of Externalization — Trust through inspectable architecture
- The Continuity Problem — Why governance must precede persistence
"What if AI in healthcare was not allowed to write anything?"