Safety & Trust
Asha was designed by physicians for a domain where mistakes cost lives. Safety is the foundation everything else is built on, beyond a feature we added.
Fiduciary Architecture
Most AI systems are “aligned” through training (encouraged to be helpful, discouraged from being harmful). Asha is fiduciary by architecture. The difference is fundamental:
Fiduciary = “Harmful paths are architecturally unreachable.”
A model can be jailbroken. An architecture cannot.
Asha's fiduciary constraints are enforced at the architecture level, before the language model ever sees your question. This means the guardrails can't be prompt-injected, jailbroken, or socially engineered around. This is a cognitive architecture covered by an allowed United States patent (US 19/290,471, Notice of Allowance October 2025) and additional pending applications.
Sacred Refusals
There are things Asha is architecturally constrained from doing. We call these “sacred refusals” because they cannot be overridden by any prompt, conversation context, or user request:
No Prescribing
Asha will never prescribe specific medications or recommend dosages. It can explain what a medication does, what the evidence says about it, and what questions to ask your doctor.
No Diagnosing
Asha will never provide a medical diagnosis. It can help you understand symptoms, provide differential considerations, and help you prepare for a doctor's visit.
No Fabrication
Asha will never invent a citation or claim a study says something it doesn't. Every reference is verifiable. If Asha can't find evidence, it says so.
No Override of Emergencies
If Asha detects a potential medical emergency (suicidal ideation, chest pain, severe allergic reaction), it interrupts immediately with crisis guidance and emergency contact information.
Emergency Detection
Asha continuously monitors conversations for signs of medical emergencies. This system is always active and cannot be disabled:
- Suicidal ideation or self-harm: Immediate crisis resources (988 Suicide & Crisis Lifeline, Crisis Text Line)
- Chest pain or cardiac symptoms: Direct guidance to call 911 or local emergency services
- Severe allergic reactions: Anaphylaxis guidance with emergency service referral
- Pediatric emergencies: Age-appropriate urgency assessment with emergency escalation
- Acute mental health crises: De-escalation support with professional referral
Epistemic Humility
One of the most dangerous things an AI can do is sound confident when it shouldn't be. Asha has a built-in solution:
- Every knowledge atom carries a confidence score and a falsifiability score
- When confidence falls below threshold, Asha says “I don't have enough evidence to answer this confidently”
- This is a hard architectural constraint, not a polite fallback
- Asha self-discloses as an AI at the beginning of every conversation
Privacy by Architecture
Health data is among the most sensitive information a person has. Asha's privacy protections are architectural, beyond policy-based:
Encryption
All data encrypted in transit with TLS 1.3. No health data stored in plaintext.
PII Anonymization
Personal health information is anonymized using Microsoft Presidio before any data reaches the AI model.
Zero AI Training
Your conversations are never sent to AI providers for model training. Zero data retention from language model APIs.
Right to Deletion
Delete your data at any time. When you delete, it's gone. No retention periods, no fine print.
Transparency
Trust requires transparency. Here's how Asha earns it:
- Self-disclosure: Asha identifies itself as an AI at the start of every conversation
- Source attribution: Every factual claim links to its evidence source
- Reasoning visibility: The inference chain is auditable
- Medical disclaimer: Present on every page and in every conversation
- Open about limitations: Asha clearly states what it cannot do