If an AI scribe drafts the progress note from a session, who owns the medical-necessity language and the liability when it's wrong?

The shift

Turning a therapy session into a structured, insurance-defensible progress note — the transcription, the clinical framing, the "medical necessity" wording payers expect — goes from scarce clinician drafting time to near-instant and abundant. Ambient AI scribes (Mentalyc, Twofold, and the wave behind them) listen or ingest a recording and return a formatted note draft in seconds. What stays with the clinician is the signature, the judgment, and the risk.

The axioms

• A defensible clinical note is skilled labor: converting a session into structured, coded, payer-ready documentation takes trained clinician time — scarce drafting.
• Writing "medical necessity" language that survives a payer audit requires fluency in the codes, thresholds, and what an auditor looks for — scarce specialist knowledge, unevenly held.
• The clinician's signature attests that a licensed human judged the session and stands behind the record — scarce accountability that a tool cannot carry.
• Judging what actually happened in the room — the risk level, the diagnostic picture, what to do next — runs on contextual reading of incomplete, emotionally loaded material with no clean rule — scarce judgment under stakes.
• The therapeutic relationship is the mechanism of change, not a delivery vehicle for information — scarce human trust.
• What is said in a session is private and controlled by consent between two people in a bounded room — scarce, contained confidentiality.
• Documentation time is a real cost that trades directly against caseload and billable hours — scarce clinician attention.

Invalid axioms

1. A defensible progress note is skilled clinician labor. Converting session content into a structured SOAP or DAP note — history, presentation, intervention, plan — is a fast first draft now, not a 10-to-15-minute writing task per client. Habit-trap: practices still treat "note-writing time" as a fixed clinical cost and still valorize documentation burden as part of the job, when the drafting is the part that just went cheap and the reviewing is what's left.
2. Producing medical-necessity language requires scarce specialist fluency. The phrasing payers want — linking symptoms to functional impairment, justifying frequency and level of care against a diagnosis — is exactly the well-documented, pattern-matched language a model reproduces reliably. Habit-trap: newer or less audit-savvy clinicians were the ones who under-documented and lost claims; a scribe now floors everyone at fluent-sounding necessity language. The risk inverts — the language is easy to generate, and the scarce act is confirming it's true of this client, not confirming the clinician can write it.

Unchanged axioms

1. The clinician's signature is an accountable human attestation, and a tool cannot carry it. A model can draft the note; it cannot be named in a licensing-board complaint, deposed in a malpractice suit, or held to a payer's clawback. When an AI-drafted note is wrong, liability lands on the clinician who signed it — the signature means "I judged this," and regulators and courts still read it that way. The drafting got abundant; the ownership did not move an inch.
2. Owning whether the medical-necessity claim is true stays a clinical judgment. Generating defensible-sounding language is not the same as the care actually being medically necessary. Whether this client at this frequency meets the threshold is a judgment about a real person; a fluent draft that overstates impairment to justify sessions is fraud whether a human or a scribe wrote it, and the accountable party is still the one who signed.
3. Judging risk and the clinical picture is a human act under stakes. Whether a client is escalating toward self-harm, whether a disclosure changes the safety plan, whether the diagnosis fits — this is judgment against ambiguous, contradictory, live material. A scribe transcribes what was said; it does not weigh what it means or catch what the client didn't say. Confidently wrong is more dangerous here than in most places a scribe gets used.
4. The therapeutic relationship is the intervention. Nothing about faster documentation changes what makes a client feel safe enough to disclose, or what makes the alliance the thing that heals. Note-drafting was never the therapy; automating it doesn't touch the work that is.

New axioms

1. Recording consent and privacy for the scribe itself is unsettled and contested. An ambient scribe means a third-party vendor is processing a recording or transcript of a therapy session — the most sensitive data category there is. Whether the client consented to being recorded and sent to an AI vendor (distinct from consenting to treatment), whether the BAA actually covers how the vendor trains or retains data, and how state two-party-consent and wiretap law apply to a clinical recording are open, and the answers vary by jurisdiction. The consent that used to be bounded by two people in a room now has a vendor in it.
2. The confidently-wrong note the clinician rubber-stamps. A scribe that mishears a name, invents a symptom the client never reported (a documented failure mode of transcription-plus-generation), attributes the therapist's words to the client, or smooths a hesitation into a clean clinical statement produces a note that reads more polished than a hand-written one — which makes it less likely to get scrutinized, not more. The polish is the hazard: it lowers the clinician's guard exactly where verification matters most.
3. Liability when an AI-drafted note misstates risk. If a scribe under-documents a suicidality disclosure, or the clinician signs a draft that omits a risk the session actually contained, and there's a bad outcome, the record now shows a note that looks thorough and was never truly reviewed line by line. Who is liable is not legally new — the clinician is — but the failure mode is: the gap between "signed" and "actually verified" is now cheap to create and hard to see after the fact, and boards, courts, and malpractice insurers haven't settled how to treat a note that was AI-drafted and human-approved.
4. The verification workload has no home in the schedule. If drafting time collapses, the residual scarce act — reading the draft against what actually happened and correcting it — is real work that doesn't feel like work, so it gets compressed. Nobody has decided whether the freed time becomes careful review, more clients, or just vanishes, and the default is that it vanishes.

Where it breaks

Practices still treat the note as clinician labor to be sped up (INVALID: drafting is now cheap) — so the pitch for a scribe is "get your evenings back," and the freed time is quietly reabsorbed as more clients or less admin, not as review capacity. But the residual scarce act moved to verifying that an abundant, polished draft is true of this client before signing (NEW), and a note that reads more fluent than a human's invites less scrutiny, not more. The signature still means "a licensed human judged this" (STILL HOLDS: accountability didn't move) — but the scribe makes it cheap to produce a note that reads like careful judgment without anyone having verified it end to end, and that gap is exactly what a board complaint or a payer audit will pull on. Whether the signature is still true in practice is the thing the workflow stops anyone from checking.

Calibration note: the confidently-wrong-transcription failure mode is on a fast-improving curve — scribe fidelity in mid-2026 is materially better than a year prior, and the drafting-quality argument may weaken. The consent, liability, and "who verified" problems are not on that curve; they are structural to putting a vendor and a probabilistic draft between the session and the signed record, and they don't get solved by a better model.

ai-axioms.vercel.app/axioms/healthcare/mental-health-scribe-ai-axioms