A prospective quality improvement study conducted at Stanford Health Care between May and July 2025, published in JAMIA Open in June 2026, surveyed 2,202 patients after outpatient visits in which an ambient AI scribe was used. The headline numbers: 70.1% of patients found the ambient AI scribe helpful. 73.6% preferred future use. 85.6% expressed interest in new technology generally.
These are strong numbers. They are about to be cited everywhere — in vendor decks, in deployment proposals, in board presentations from health system executives looking for evidence that ambient AI documentation is patient-acceptable. The Abridge class action defense team may cite them. The lobbyists opposing all-party consent legislation will cite them.
The problem is not that the numbers are wrong. The problem is what they measured — and what they explicitly did not.
The three variables that were excluded
The authors of the Shah et al. study are transparent about this. In the discussion section, they state directly: “This study also did not specifically assess patient perceptions related to communication with their clinician, privacy, trust, or perceived accuracy.”
Privacy. Trust. Communication.
These are not peripheral variables. They are the three dimensions that determine whether patient acceptance of ambient AI documentation is durable, informed, or genuinely consented to. A patient who finds the tool helpful because the doctor made more eye contact during the visit has not consented to their audio being transmitted to a third-party vendor, processed outside the clinical setting, or used to improve the vendor’s model. Helpfulness and consent are not the same construct.
The study measured what patients noticed during the visit. It did not measure what they understood about what happened to their data afterward.
The demographic problem
The study’s patient population deserves the same scrutiny as its measurement scope. 59.8% of respondents were aged 70 or older. The study was conducted at Stanford Health Care — an academic medical center serving a predominantly educated, English-speaking, technologically engaged patient population in one of the most tech-saturated regions in the country.
This is not the US patient population. It is not the Medicaid population. It is not the rural population. It is not the population served by safety-net hospitals where ambient AI adoption is now accelerating. The authors acknowledge this limitation directly.
There were statistically significant differences in perceived helpfulness across race — Asian respondents rated the tool as helpful at 86%, White respondents at 74%, Black respondents at 78%. The study does not examine why. It flags these differences and defers to future qualitative work. That future qualitative work does not yet exist. It should before the 70% number becomes the industry’s answer to every consent question.
The consent model that was used
Clinicians in the Shah et al. study were required to notify patients that the visit would be recorded before activating the ambient AI scribe tool. Patients could decline. That is the entire consent architecture — the notification model: a verbal statement before recording begins, with no written documentation, no disclosure of where the audio goes, and no record of the conversation about consent itself.
At Stanford Health Care, this consent model produced 70.1% helpfulness and 73.6% preference for future use. It also, in a separate organization using the same ambient AI documentation vendor, produced a class action lawsuit alleging 100,000+ patients were recorded without clear notice that their medical conversations would be transmitted outside the clinical setting or processed through third-party systems.
The Abridge lawsuit — filed November 26, 2025 — is the first major legal challenge to ambient AI clinical documentation in the US. The case details are instructive.
Case No. 25CU063632C · Superior Court of California, County of San Diego · Filed November 26, 2025 · Hon. Carolyn M. Caietti presiding · Counterpoint Legal for plaintiff · Status: Open
Defendants: Sharp HealthCare; Sharp Rees-Stealy Medical Group; SharpCare Medical Group; Sharp Community Medical Group; Does 1–50
What happened: Jose Saucedo attended a routine physical exam at Sharp Rees-Stealy Medical Group in July 2025. Audio of the encounter was captured via the clinician’s microphone-enabled device using Abridge’s ambient documentation platform, transmitted to Abridge’s cloud servers, and used to generate a clinical note. Saucedo was neither informed nor asked to consent. Sharp launched the Abridge deployment in April 2025.
The false consent record: Saucedo discovered the recording by reading his patient portal notes. The AI-generated chart language stated he had been “advised” the visit was being recorded and had “consented.” The complaint calls that language false — inserted by the system, not reflecting any actual disclosure. When Saucedo requested deletion, Sharp told him the vendor retains audio for approximately 30 days and could not delete it promptly. Sharp offered to modify the AI-generated note instead.
Causes of action: (1) Violation of California Invasion of Privacy Act, Cal. Penal Code §§ 632 & 637.2 — all-party consent required before recording; (2) Violation of California Confidentiality of Medical Information Act — written authorization required before sharing PHI with third parties; (3) Invasion of Privacy — Intrusion Upon Seclusion; (4) Negligent Misrepresentation. Jury trial demanded.
Class and exposure: Plaintiff seeks class certification for any California resident with a Sharp HealthCare visit audio-recorded without consent on or after April 1, 2025. Estimated class: 100,000+ patients. CIPA provides up to $5,000 per violation, per recording.
Court record: San Diego Superior Court → Complaint text via Trellis: trellis.law →
The distance between the Stanford result and the Abridge lawsuit is not a technology difference. It is an implementation difference. The same notification-only consent model produces radically different outcomes depending on whether it is executed with institutional care or institutional negligence. A governance framework designed around the notification floor produces exactly this distribution — some acceptable, some fraudulent, all legally exposed in the 11 states that currently require all-party consent before recording.
What the broader literature shows
The Shah et al. paper sits within a rapidly expanding evidence base. A real-world time-motion study published in JMIR Medical Informatics in March 2026 found that ambient scribes reallocated clinician effort toward patient interaction — a genuine benefit — but also flagged equity gaps in multilingual settings that the Stanford study does not address. A barriers and scaling analysis published in npj Digital Medicine the same month specifically called out the need for explicit informed consent in telehealth ambient use, where patients may be in home environments with other people present.
The first randomized controlled trial of ambient AI scribes, published in NEJM AI in 2026, provides the strongest methodological anchor in the field. Its findings on clinician burden are significant. Patient consent infrastructure remains out of scope — which is precisely the gap this post is flagging.
The NHS issued a DPIA template for ambient scribes in March 2026. The UK has a governance framework. The US has KFF Health News asking patients whether they can opt out.
What rigorous acceptance would actually look like
The Shah et al. findings are a starting point, not a conclusion. Rigorous acceptance — the kind that holds up to a class action, a regulatory investigation, or a patient who later learns what happened to their data — requires four things the current notification model does not provide.
Documented consent with opt-out tracking. Not a verbal notification but a recorded consent event that is part of the patient record, linked to the encounter, and auditable. The 18.9% of patients in the Shah et al. study who said they would not want the tool used again need to show up in a consent registry — not just in a post-visit survey.
Disclosure of data flows. Patients need to know that their audio is transmitted to a third-party vendor, that de-identified data may be used to improve the model, and that audio retention policies vary by vendor. Helpfulness during the visit is not a proxy for this understanding.
Privacy and trust measurement. The Shah et al. study explicitly deferred privacy and trust to future research. That future research needs to happen — and to happen before the tool is deployed at scale in populations with lower baseline trust in healthcare institutions.
Equity-aware implementation. The statistically significant racial differences in the Shah et al. data are not explained. They should not be papered over with the aggregate 70% figure in deployment proposals that do not look like Stanford Health Care.
What the Consent Gap series established
Part 1 of the Consent Gap series documented that no federal law requires specific AI disclosure to patients before ambient documentation is used. HIPAA governs data handling after the fact. The notification model is the current legal floor — and it is being used as the ceiling.
Part 2 documented that the research — the Mello/JAMA framework, the Michigan TIERRA patient survey — consistently finds patients want active, informed consent, not opt-out checkboxes. The Shah et al. paper does not contradict this. It measures helpfulness, not consent quality. Patients who found the tool helpful in a verbal-notification-only model are not the same as patients who gave informed, documented, revocable consent.
Part 3 posed the question no one has answered: when the ambient scribe is running across every encounter, who holds the consent lineage — the trail linking the original consent event to the data flows, model versions, and inference pathways that followed?
The Abridge lawsuit provides the answer when that question goes unasked.
Ambient AI scribe acceptance at 70% under notification-only consent is a starting point. The three things that study didn’t measure — privacy, trust, and communication — are the three things that determine whether that acceptance holds when patients find out what happened to their data.
The industry needs this number to mean more than it means. The governance architecture that makes ambient AI documentation genuinely defensible is the same one the number cannot yet prove exists. Build it before the 70% becomes a liability.
