Behavioral Health AI Documentation: HIPAA and 42 CFR Part 2

AI documentation tools are spreading through behavioral health — therapy session notes, SUD treatment records, intake assessments — and this is the specialty where general-purpose compliance thinking fails, because behavioral health runs on two stacked regimes: HIPAA's general framework plus 42 CFR Part 2's consent-driven rules for substance-use records, with psychotherapy-note protections layered over both. A scribe that satisfies HIPAA for a cardiology practice can be non-compliant for a therapy practice. This page maps the regime stack, assigns duties across practice and vendor, and turns Part 2's segmentation demand into a verification protocol.

Scope: What Part 2 Adds for Therapy and SUD Records

Behavioral health documentation carries two stacked regimes: HIPAA's general safeguards plus 42 CFR Part 2 for substance-use-disorder records, which requires specific patient consent for disclosures and record segmentation — and therapy notes add the psychotherapy-notes overlay whose heightened protection has near-identical definitions under both frameworks.

The Final Rule updated Part 2's mechanics — single consent for future treatment-payment-operations uses, breach notification now applying to Part 2 records, closer alignment with HIPAA covered entities — but the consent-driven core persists: Part 2 records disclose per specific patient consent, not broad permission. Compliance guidance notes the analogy that makes the stakes legible: Part 2-protected records function like HIPAA psychotherapy notes, the most protected category HIPAA defines. Whether a given record is Part 2-covered depends on the program's SUD treatment context, which makes classification a compliance-review task — this table supplies the inputs.

Responsibility for AI Scribes in This Regime

The practice owns consent capture, segmentation policy, and the risk analysis; the scribe vendor is a business associate whose BAA must also respect Part 2's consent-driven disclosure limits — meaning the vendor contract, not just the tool, must refuse disclosures the consent does not authorize.

The vendor row is where behavioral-health deployments most often go wrong: a BAA written for general medicine makes business-associate promises that Part 2 exceeds, because Part 2's refusal duty — do not disclose what consent does not authorize — is a contractual behavior, not a security control. Legal review of the vendor contract's Part 2 clauses is a prerequisite, not a formality, and the chain check extends to every subprocessor that can reach the records.

Consent Segmentation as an Architecture Requirement

Part 2's consent model demands that records be segregable — stored, exported, and disclosed in consent-shaped slices — so the AI tool must support field- and record-level segmentation, consent-scoped exports, and no-training terms, verified by test rather than asserted by badge.

• The segregability test: can the tool's output be stored and exported such that a Part 2 record moves without dragging non-consented content with it — and vice versa?
• Consent-scoped export: when a disclosure request arrives, the export contains exactly what the consent authorizes, with the consent record attached.
• No-training terms: contractual confirmation that session content never improves vendor models — therapy notes are the last data anyone wants memorized.
• Verification, not assertion: run the protocol on test records with a synthetic consent matrix, before the tool touches real sessions.

Vendor guidance for this specialty converges on exactly these checks — consent, segmentation, SUD-note handling — and the test protocol is what converts them from badge-reading to evidence. Design the synthetic consent matrix to cover the awkward cases: partial consents, revoked consents, and disclosures that would be routine under HIPAA but unauthorized under Part 2.

Residual Risks Specific to Behavioral Health

Record the residuals: re-identification risk is highest in small-population therapy notes, session audio carries the same Part 2 status as notes, model outputs may embed clinical detail into general records, and consent revocation must propagate through every stored derivative — each with an owner and cadence.

The revocation row deserves emphasis: Part 2 consents can be revoked, and a revocation that clears the notes but not the audio, the drafts, or the vendor's transcripts is incomplete in exactly the way audits find. Residual acceptance needs clinical and legal input — this register frames that review. Where the risk analysis concludes behavioral-health workloads belong inside the organization's own boundary, dedicated environments such as OneSource Cloud's healthcare AI infrastructure are one self-hosted path for the pipeline, under the same Part 2 checks as any vendor.

FAQ

What is the difference between HIPAA and 42 CFR Part 2 for AI notes?

Part 2 is consent-driven where HIPAA is permission-driven: HIPAA allows treatment-payment-operations disclosures broadly, while Part 2 SUD records disclose only per specific patient consent and must be segmentable — so an AI tool that satisfies HIPAA generally can still fail Part 2 on consent and segmentation.

Did the Final Rule make Part 2 the same as HIPAA?

It aligned mechanics — single consent for TPO, breach notification — but the consent-driven core for Part 2 records persists, so the practical answer is closer alignment, not equivalence; behavioral-health deployments still need Part 2-specific consent and segmentation.

Can we use a general medical AI scribe for therapy sessions?

Only if it passes the Part 2 checks general tools rarely built: consent-scoped segmentation, disclosure refusal beyond the consent, no-training terms, and revocation propagation — otherwise the tool is compliant for medicine and non-compliant for your therapy notes specifically.