California health system secretly recorded patients with AI scribe, lawsuit claims

What happened

In November 2024, a patient filed a proposed class-action lawsuit against Sharp HealthCare in California, claiming the health system secretly recorded doctor-patient conversations using AI medical scribe technology ^[1]. The complaint alleges Sharp did not tell patients their words were being recorded, transcribed and stored by an AI system, which the plaintiff says violates California's Invasion of Privacy Act ^[1].

The AI Incident Database, maintained by the Responsible AI Collaborative, catalogues the case as incident 1754 ^[1]. According to the database entry, the lawsuit claims Sharp used an AI scribe from a company called Abridge without informing patients or obtaining their consent ^[1].

AI medical scribes listen to consultations, transcribe them and draft clinical notes. The technology is spreading fast in healthcare because it saves doctors time. The privacy question is straightforward: if a system records what you say in a medical consultation, you ought to know about it.

The evidence question

Think of it like the black box on an aeroplane. After an incident, investigators want the recording. The first question is whether it exists. The second is whether it's complete and unchanged. The third is who has listened to it.

In this case, the questions are the same. What exactly did the AI record during each consultation? Who accessed the recordings, and when? Were the recordings stored securely, and can the hospital show they haven't been altered since they were made?

Those are the questions a court, a regulator or the patients themselves might ask, and the answers depend on records the hospital keeps. If the only record of access is a log the hospital's own administrators can edit, it's harder to answer with confidence.

What the rules say

In Sweden, healthcare providers must keep patient records and document who has accessed them, under the Patient Data Act (Patientdatalagen) ^[2]. The duty applies to all forms of record, whether written by a doctor, typed by a secretary or drafted by an AI.

The EU's AI Act doesn't yet apply to this case, but it will matter soon. High-risk AI systems must technically allow for the automatic recording of events over their lifetime, under Article 12 ^[3]. For AI that falls under Annex I of the Act, including certain medical devices, that requirement applies from 2 August 2028 ^[4].

Article 12 says what to log. It doesn't say how to prove the logs are intact, but the question is real: when someone asks what the AI recorded, and who saw it, the organisation's answer depends on records it can stand behind.

The Medical Devices Regulation requires manufacturers of medical software to monitor what happens after a product reaches the market and to report serious incidents ^[5]. An allegation that patient data was recorded without consent, or accessed by unauthorised parties, would be exactly the sort of incident that triggers those duties.

A sealed record of what was recorded

This is where a digital witness fits. The thing to seal is each scribe session: what was recorded, when it was captured, who requested the transcript and which version of the AI produced it. Sealing takes a fingerprint of that record, not the recording itself, so the content never leaves the hospital's systems.

If a patient or a court later asks what was recorded on a given day, the hospital can produce the record and show it hasn't changed since it was sealed. Anyone the hospital shares it with can check the fingerprint and see for themselves.

A digital witness doesn't show that consent was obtained, or that the recording was lawful. It shows that the record of what happened is the one written at the time, unchanged since. When the question is whether access logs or session records have been quietly edited, that distinction matters.

How it would work

1. Each time the AI scribe completes a session, the system seals a record containing the patient reference, the date, the length of the recording, the clinician and who accessed it.
2. The hospital keeps the recordings and the clinical notes in its own systems, as it does now. Only a fingerprint of each record goes to WitniumChain.
3. If the record is later questioned, the hospital calculates the fingerprint again. If it matches, nothing has changed. If it doesn't, someone has altered it.

The first 2,000 witnesses a month are free, through the REST API, the SDK or the MCP server. Read the docs to start.

What this means for you

If you run or build healthcare technology that records consultations, drafts notes or uses AI to assist clinicians, here are the questions to ask now:

• Could you show exactly what your AI recorded during a specific consultation six months ago?
• Can you prove who accessed that recording, and when?
• If a patient or a regulator questioned whether an access log had been altered, what would you show them?
• Do your systems technically allow for the automatic recording of access events, in a way that meets the duties you already have under data protection and healthcare law?

The Sharp case is one lawsuit in California, not a ruling. But the questions it raises apply wherever AI scribes are used, because the duty to protect patient data and document access already exists.

What to watch

The case itself, Sharp HealthCare, is in San Diego Superior Court and had not reached a decision as of early 2025 ^[1]. Whether it settles, is dismissed or goes to trial, it won't be the last dispute about what an AI medical scribe recorded.

For healthcare providers in the EU, the Medical Devices Regulation already applies ^[5]. The AI Act's high-risk requirements for Annex I systems, including certain medical software, apply from 2 August 2028 ^[4]. The Article 12 logging requirement is part of that package ^[3].

In Sweden, the Patient Data Act has required documentation of record access since 2008 ^[2]. AI scribes don't change that duty; they make it more urgent, because the question of what was recorded and who saw it can now come up in ways it didn't before.

The technology is useful, and it's spreading. The question is whether the organisations using it can show, later, what their systems actually did.