Can Therapy Transcripts Be Subpoenaed? What Kamrass v. AdventHealth Means for Every Digital Mental Health User

Yes — and it already happened. On April 28, 2026, Proof News published an investigation by Annie Gilbertson revealing that Jennifer Kamrass, a nurse practitioner at AdventHealth, had every word she typed to her Talkspace therapist subpoenaed and produced in court — by the same employer she had just sued for pregnancy discrimination. Under HIPAA's court-order exception at 45 CFR 164.512, that disclosure was lawful. Which is the problem.

For anyone who assumed a digital therapy app was as confidential as a couch in a psychologist's office, the Kamrass case is the moment the illusion cracked. The rest of this article walks through exactly how it happened, why HIPAA didn't stop it, and why the only durable fix is architectural: if a verbatim transcript never leaves the device, there is no vendor-side record for opposing counsel to subpoena.

What Actually Happened in Kamrass v. AdventHealth

Jennifer Kamrass was a nurse practitioner at AdventHealth. According to the Proof News investigation, she used Talkspace — offered through her employer's benefits program — to process anxiety about her marriage, finances, and self-esteem during a period that included pregnancy and job loss. She was terminated while nearly nine months pregnant and filed a pregnancy-discrimination claim.

Her therapist agreed to testify on her behalf. In response, AdventHealth's counsel subpoenaed the underlying Talkspace records — the full message history between Kamrass and her therapist. As the Psychotherapy Action Network wrote in its clinical analysis, digital platforms that record every exchange do not replicate a therapy session — they transform it into a document, and documents can be subpoenaed. A federal judge ultimately ruled for AdventHealth on the merits.

Her lawyer, Peter Andreone, told Proof News: "You've now got written evidence of everything discussed and in a normal therapy session that wouldn't be true. They turned around and used it against her."

Why HIPAA Didn't Protect Her

The instinctive reaction is: how is that legal? The uncomfortable answer is that HIPAA is a data-disclosure statute, not a discovery-blocking one. As Captain Compliance's breakdown of the case explains, three pieces of HIPAA's architecture combined to allow the transcript into court:

• The court-order exception. 45 CFR 164.512(e) explicitly permits covered entities and business associates to disclose PHI in response to a court order or subpoena, subject to notice requirements. This is not a loophole — it is a deliberate carve-out.
• Psychotherapy-notes carve-out doesn't reach the full transcript. HIPAA's stronger protection for "psychotherapy notes" (which require separate patient authorization) applies only to the clinician's personal process notes kept separate from the medical record. As the ComplianceHub analysis of digital therapy compliance in 2026 notes, Kamrass's transcripts were classified as the treatment record itself, not psychotherapy notes.
• Digital-therapy platforms may not even be HIPAA-covered. Many operate under state consumer-privacy law and FTC Health Breach Notification rules rather than as full HIPAA business associates.

Even if every HIPAA control worked exactly as designed, the outcome would not change. HIPAA was written to prevent unauthorized disclosure — not to shield litigants from discovery.

The 140-Million-Message Training Set

The Kamrass subpoena is only half the story. The other half is what the platform is doing with everyone else's transcripts. AI Governance for HR reported that Talkspace has publicly disclosed to investors that it is building large language models trained on what it describes as the industry's largest behavioral-health dataset — roughly 140 million message exchanges over twelve years, now powering an AI product.

Talkspace says the sessions are de-identified. Researchers say that is not a meaningful guarantee. As the follow-up Proof News investigation into AI scribes and therapists documented, word patterns, emotional cadence, and the therapeutic arc of an individual's disclosures over time are identifying signals even without a name attached. The same investigation found one telehealth platform's supposedly de-identified transcripts had been exposed in court.

In March 2026, Universal Health Services agreed to acquire Talkspace for $835 million. The dataset traveled with the company. The terms under which patients originally consented to share their most private thoughts did not automatically transfer to a new owner.

Cloud Therapy vs. On-Device: A Legal-Exposure Comparison

The Kamrass case is not just about Talkspace. It is a preview of how any cloud-based therapy tool creates the same exposure surface. Here is the architectural comparison:

The Parallel Case: Ambient AI Scribes and Washington v. Sutter Health

The Kamrass case is the chat-transcript version of a story that is now playing out for spoken clinical conversations too. On April 8, 2026, three plaintiffs filed Washington v. Sutter Health in the U.S. District Court for the Northern District of California, alleging that Sutter and MemorialCare used Abridge's ambient AI scribe to record patient-clinician conversations without informed consent.

As Alston & Bird's privacy team wrote in their analysis, the complaint focuses on the recording itself — the legal violation is said to occur "at the moment of interception," when live communications are captured and transmitted to third-party servers. The Federal Wiretap Act, the California Invasion of Privacy Act (CIPA), and the California Confidentiality of Medical Information Act (CMIA) are all in play. HIPAA Journal's reporting on the case noted the plaintiffs allege they had no idea the AI platform was recording them.

Different modality, identical architectural problem: audio or text of a clinical conversation is transmitted to a vendor server, where it becomes a permanent, discoverable, breachable, and potentially trainable record. We covered the broader wave in our Abridge AI & Sutter Health explainer.

The AI-Notetaker Wiretap Wave Is Not Just a Healthcare Problem

Zoom out and the pattern is the same across categories. A federal class action, Chamberlain v. Granola, was filed July 30, 2026 in the Northern District of California. According to the National Law Review's analysis by Robinson & Cole, the complaint alleges Granola's AI notetaker recorded a virtual meeting participant without notice and used meeting contents by default for commercial purposes including AI training, unless the Granola user turned that setting off.

Meanwhile, HR Executive reported that Chamberlain v. Granola sits in the same district as the consolidated Otter.AI Privacy Litigation, where oral argument on Otter's motion to dismiss was set for August 2026. The plaintiffs' theory in every one of these cases — Otter, Fireflies, Granola, Abridge, Talkspace-adjacent — depends on the same fact: audio or text was transmitted to a third-party server. Remove that transmission and the theory has no hook. Our deep dive on Chamberlain v. Granola unpacks the consent-design failure in detail.

What Regulators Are Now Saying About AI and Mental Health Data

The regulatory backdrop is tightening at the same moment the litigation is landing. The EU AI Act's transparency obligations under Article 50 took effect in August 2026, requiring users to be informed when they interact with an AI system. And under GDPR Article 9, mental-health data is a "special category" that requires an additional lawful basis beyond ordinary consent — a bar that is very difficult to clear when a platform is using session transcripts to train an AI model shipped to a downstream buyer.

In the U.S., the California Attorney General's CCPA enforcement guidance treats mental-health data as sensitive personal information subject to stricter use limits. And as MyPrivacy's 2026 digital-therapy wrap-up documents, the FTC's Health Breach Notification Rule has been applied to non-HIPAA digital health apps repeatedly in 2025–2026.

What Clinicians Should Do This Week

If you are a therapist, psychiatrist, coach, or clinical supervisor using any AI-assisted documentation tool, here is a copy-into-your-policy checklist derived from the Kamrass and Sutter fact patterns:

1. Map where the verbatim record lives. If your vendor stores audio or transcripts on its servers, assume it is discoverable under 45 CFR 164.512(e).
2. Confirm training-data policy in writing. "We don't sell your data" is not the same as "we don't train on your data." Ask directly whether de-identified transcripts feed models, and whether that changes on acquisition.
3. Set the shortest retention window your state and payer rules allow. Talkspace retains transcripts as 10-year medical records; you may be able to reduce this materially.
4. Disclose to clients before recording. All-party consent is not just a California problem — the CIPA analysis of Otter.ai litigation notes wiretap statutes exist in eleven states.
5. Prefer on-device transcription for sessions with heightened sensitivity. Custody battles, workplace disputes, and immigration cases are exactly the fact patterns where transcripts get subpoenaed.

How Basil AI Solves This: The On-Device Architecture

Basil AI runs on Apple's on-device Speech framework and the Apple Neural Engine. Session audio is captured, transcribed, and summarized entirely on the clinician's iPhone, iPad, or Mac. Nothing is transmitted to a Basil server — because Basil does not run a transcription server. That architecture maps directly onto Apple's public privacy commitments, which state that on-device processing keeps personal information on the device.

For the Kamrass-style discovery risk specifically, that has three consequences:

• No vendor subpoena target. Opposing counsel cannot serve a subpoena on Basil for your patient's transcripts, because Basil does not hold them. Discovery follows the therapist's normal record-custody rules — the same rules that governed paper progress notes for decades.
• No training pipeline. Your session transcripts cannot feed a large behavioral-health dataset, because they never leave the device. There is no acquirer scenario in which "the dataset traveled with the company."
• No CIPA/CMIA/Wiretap Act interception element. The plaintiffs' theory in Washington v. Sutter Health depends on audio being transmitted to a third-party server. Remove the transmission and the theory has no hook.

None of this is a compliance guarantee — HIPAA, CMIA, state licensing law, and payer rules still apply, and the clinician remains the covered entity. But it is a materially smaller attack surface. For a deeper technical walkthrough, see our on-device therapy transcription architecture piece and our comparison of how AI meeting notes are being treated in discovery.

The Bottom Line

Kamrass v. AdventHealth did not create new law. It exposed old law working exactly as written on a new fact pattern that most patients — and most clinicians — never imagined. Every verbatim transcript sitting on a vendor server is a document. Every document is discoverable. Every acquired company brings its dataset with it. Every large behavioral-health corpus is a target for training, breach, or subpoena.

The only architectural intervention that changes the outcome is the one that removes the vendor from the record entirely. That is what on-device transcription does. That is what Basil AI does. And it is why, for the most private conversations most people ever have, the cloud was never the right place for them to live.