AI scribe data, privacy, and model training
AI scribe privacy in psychiatry is dominated by three questions: is audio retained after note generation, is customer data used to train models, and which subprocessors touch the record. All three have vendor-controlled defaults that clinicians rarely see in the sales cycle. The answers materially change what psychiatric practices should be comfortable putting through the scribe — and every question can and should be resolved in writing before the first real patient visit.
Audio retention: the highest-risk default
The safest default is that audio is transcribed, used to produce the note, and immediately discarded. Some vendors do this by default; others retain audio for a rolling window (7, 30, or 90 days) for quality review or model training. Retained audio is the highest-risk artifact in the pipeline — it contains raw patient speech, including identifiers that may not appear in the finished note.
Ask each vendor: what is the audio-retention default on my tier? Can I turn it off? If I turn it off, does that limit any product feature I depend on?
Model training on customer data
The industry has been moving toward opt-out or off-by-default in 2025-2026, but defaults still vary. Get the following in writing: is my note text used to train models? Is my audio used to train models? Is any de-identified or aggregated data used to train models, and how is de-identification verified? What is the process to opt out and does it apply retroactively?
For psychiatry specifically, aggregated patterns of psychiatric assessment and language are more identifying than they look. Insist on opt-out at minimum, and prefer vendors that opt customers out by default.
Subprocessor stack
Every AI scribe uses at least one large-language-model subprocessor — usually OpenAI, Anthropic, or Google — plus cloud infrastructure (AWS, GCP, or Azure) and often a separate speech-to-text provider. Nabla and Twofold publish full subprocessor lists; other vendors provide the list on request. If the list changes materially, the vendor should notify you in advance; confirm that obligation is in the BAA.
Downstream BAAs
Each subprocessor that touches PHI must be covered by a downstream BAA. Confirm the primary vendor holds these — you should not be signing separate BAAs with each subprocessor.
Geographic scope
Some subprocessors process data in specific regions; if your compliance posture requires U.S.-only processing, confirm the primary vendor enforces that.
Deletion, portability, and offboarding
When you leave a vendor, you should be able to export your notes in a machine-readable format and confirm deletion of both notes and audio from vendor systems (and subprocessor systems) within a documented window. Sixty to ninety days is a reasonable deletion SLA; anything longer is worth negotiating down.
Privacy verification checklist
- Audio retention default and opt-out mechanism
- Note-text and audio use for model training, and opt-out mechanism
- Full current subprocessor list, in writing
- Downstream BAA coverage confirmed for every subprocessor
- Data-region enforcement (if applicable)
- Deletion SLA and export format at offboarding
- Breach notification obligation within 60 days
Frequently asked
- Do AI scribes train on my patient audio?
- Some do by default. Insist on written confirmation of your tier's setting, and prefer vendors that opt customers out by default.
- Is my note text used to improve the model?
- Ask specifically. Some vendors distinguish between training on note text (higher risk) and quality review of aggregated performance metrics (lower risk); the two are not the same.
- Can I get a zero-retention mode?
- Increasingly yes, on paid tiers. Zero-retention modes discard audio immediately and either discard the transcript or store it only in the customer environment.
- Who owns the notes the scribe generates?
- The clinician and practice, per standard vendor terms in this market. Confirm the terms explicitly grant no other use rights to the vendor beyond providing the service.
Scribes referenced in this guide
Related guides
- HIPAA and AI scribes: what psychiatrists must verify
The specific HIPAA, BAA, encryption, audio-retention, and 42 CFR Part 2 checks a psychiatrist should run before using any AI scribe with real patients.
- AI scribes for addiction and SUD psychiatry
How AI scribes handle SUD psychiatric documentation under 42 CFR Part 2 — vendor readiness, MAT program notes, and workflow safeguards.