How MyPsyche uses AI
Last updated: 10 June 2026 · This page is our AI transparency notice under Regulation (EU) 2024/1689 (the EU AI Act) and our own commitments.
The one guarantee that matters
A human is always in the loop. AI in MyPsyche drafts and suggests; it never decides. Nothing — no conceptualization text, no session summary, no generated document — enters the clinical record until the treating therapist reviews and approves it. AI-drafted content is labelled as such in the product, and every draft cites the client evidence it was based on (session notes, questionnaire findings) so the therapist can verify each claim against the source.
What is AI, and what is not
- Not AI: questionnaire scoring (YSQ-L3, SMI) uses the instruments' standardized scoring keys — deterministic arithmetic, not machine learning.
- AI — drafting: case-conceptualization drafts, session-note summaries and document generation use Anthropic's Claude models via API.
- AI — transcription: session audio is transcribed and diarized by Speechmatics.
- AI — retrieval: semantic search over a therapist's own material uses Voyage AI embeddings.
Where your data goes
| Provider | Used for | Training on your data | Retention |
|---|---|---|---|
| Anthropic (Claude) | Drafting, summaries, document generation | No — contractually excluded for API data | Short fixed API log window (days), then deleted |
| Speechmatics | Transcription + speaker diarization | No | Auto-deleted after 7 days |
| Voyage AI | Embeddings for retrieval | No (training opt-out enabled) | Zero-retention setting |
Data sent to these providers is limited to what the feature needs. Generated documents use client-side tokenization so real names do not leave your browser; we are extending identifier minimization across the other AI flows.
Known limitations
- Language models can produce plausible but wrong statements. That is why drafts carry confidence levels, caveat flags and source citations — and why your review is mandatory, not optional.
- Transcription quality varies with audio quality, accents and cross-talk; always check speaker attribution before generating summaries.
- AI drafts reflect the evidence available; missing or outdated notes produce weaker drafts. The product gates drafting and says so when evidence is insufficient rather than inventing content.
What MyPsyche's AI does not do
- It does not diagnose, assess, or classify clients.
- It does not make or recommend treatment decisions.
- It does not infer emotions from audio, and we will not enable emotion-recognition features.
- It does not act autonomously in the client portal; clients are never exposed to unreviewed AI content.
Questions
Write to [CONTACT EMAIL]. We treat AI-transparency questions from clinicians and their clients as support priority.