DULUTH, MN MANILA, PH
Say hello

Case Study 2026

Ambient AI scribes

Not usingClinical AI

Summary. The pitch is real — documentation time drops, burnout improves, throughput goes up. The trade-offs are also real — patient consent is reduced to a registration-form checkbox, legal accountability remains with the clinician while cognitive authorship moves to the model, and the encounter itself changes in ways patients notice. I am declining to use these products at the bedside until per-encounter consent, on-prem processing, validated accuracy on my actual patient population, and clear institutional governance all exist together.

For the long-form argument, see the note The case against ambient AI scribes at the bedside. This page is the operational summary I share when colleagues ask “why aren’t you using it yet.”

What the technology actually is

The transcription layer is OpenAI’s Whisper or a Whisper-derivative, often packaged with a large language model that turns the transcript into a SOAP-shaped note. The novel work in a commercial “ambient scribe” product is not the microphone — it is the LLM that summarizes, the consent flow that wraps the recording, the storage policy on the audio, and the workflow that decides who reviews the draft. When marketing collapses these four very different questions into one slogan (“Is AI ready for medicine?”), the answer is misleading by construction.

Where I have landed

ConcernStatus today
Per-encounter, opt-in patient consentUsually a registration checkbox at best — insufficient
On-device / on-prem processingVendor-cloud is the default; transparency on retention varies widely
Validation on real populations (elderly, multilingual, chronically ill)Pilot data skews young, English-speaking, ambulatory
Institutional governance of draft review and disagreement trackingInconsistent across health systems
Cognitive authorship vs legal authorship gapUnresolved and being normalized at scale

What would change my mind

Per-encounter consent that is more than a checkbox. Local or on-prem processing with no third-party retention by default. Published validation on populations resembling mine. A clear governance policy at my institution about who reviews drafted notes, who is responsible for errors, and how disagreements between the model and the clinician are tracked.

Why this is a public position

If I am wrong about this in five years I will say so on this page. The point of writing it down now is to be specific about what would change my mind, so that future-me has something to revise against.


— Jeremy Tabernero, MD · More case studies · Get in touch