Ethics and safety
Safe AI Therapy: Guardrails, Ethics & Human Oversight
Safe AI therapy is mental health support where AI is used only for non-clinical tasks — journaling, mood tracking, psychoeducation, and matching — under clear guardrails, transparent labelling, and full human clinical oversight. MySafeTherapy is designed to this standard: no AI diagnosis, no AI clinical decisions, no hidden bots pretending to be therapists, UK GDPR compliance, and always-on routing to human clinicians or crisis services when risk is detected.
Key points
- Every AI feature is clearly labelled as AI — no impersonation of clinicians.
- AI never diagnoses, prescribes, or makes clinical decisions.
- Crisis language is detected and users are routed to human support and 999/Samaritans.
- UK GDPR compliant, ICO registered, data encrypted at rest and in transit.
- Journal entries are not used to train third-party AI models.
- All therapy is delivered by BACP, UKCP, or HCPC registered clinicians.
What 'safe' means at MySafeTherapy
- Transparency — every AI surface is labelled and describes its purpose.
- Scope — AI is limited to journaling, reflection, mood tracking, booking help, and psychoeducation.
- Guardrails — models are constrained by prompts that block clinical advice and unsafe content.
- Human-in-the-loop — clinicians can review shared reflections, and safeguarding concerns escalate to humans.
- Data minimisation — we collect only what's needed; users can export or delete their data.
- No dark patterns — no manipulation, no dependency loops, no fake empathy claims.
The five safety layers
- 1. Consent and clear labelling — users always know they're interacting with AI.
- 2. Content filters — prompts and outputs are filtered for unsafe medical or crisis content.
- 3. Crisis routing — trigger words open crisis-support pathways immediately.
- 4. Clinical oversight — clinicians can view shared journal entries and workbook responses.
- 5. Independent audits — regular reviews of AI outputs, privacy, and safeguarding processes.
Red flags in unsafe AI therapy
- AI 'therapist' personas that don't disclose they are AI.
- Diagnostic claims from AI ('you have anxiety disorder').
- No crisis routing or generic 'please seek help' redirects only.
- Unclear data policies or use of chat data to train foundation models.
- No route to a real human clinician.