Services

About

Process

Writing

Contact

Book a meeting

Ambient AI Scribes: Effortless is the Risk

Ambient AI scribes are giving clinicians their attention back. The question is what disappears along with the paperwork — and who's left accountable when it does.

by Kelly Smith • 3 August 2026

I recently saw one of my physicians and noticed she was paying closer attention to me than she had at my previous visit. I assumed she was just having a better day.

It turned out she was using an ambient clinical scribe — an AI tool that listens to the appointment and drafts the clinical notes for her. From where I sat, she was simply more present. The visit was no longer than usual, but it felt like she had more time for me. From her side of the exam table, she was glad to be focusing on her patient instead of her keyboard.

Our shared experience mirrors what the research is finding. A study in JAMA Network Open looked at more than 250 providers across six health systems running AI-scribe pilots and found self-reported burnout fell from roughly 52% to 39%. Participants also reported lower cognitive burden, less after-hours documentation, and a better ability to stay present with patients.

Neda Laiteerapong, MD, a UChicago Medicine internist who uses AI scribes in her daily practice, describes the change this way:

“We're more focused throughout the day and less mentally exhausted, leaving more energy and compassion to dedicate to improving patients' quality of life, even through seemingly small changes like ordering lab tests further ahead of a clinical visit.”

A separate study in NEJM Catalyst reported that most physicians had a highly positive experience with AI scribes, and that every patient surveyed rated the impact on their visit as positive or neutral.

So with the research pointing to real gains in clinician well-being — and patients responding warmly — is there anything to worry about?

There is. Four things, in fact. And none of them are reasons to abandon the tools. They're reasons to design them more carefully.

1. The time physicians "get back" is smaller than advertised — and it may not be theirs to keep

Marketing for ambient scribes leans hard on the promise of reclaimed time. So how much time actually comes back?

Less than you'd think. A 2026 JAMA multisite study of clinician time expenditure and visit quantity found the savings were modest: about 16 fewer minutes of documentation and 13 fewer minutes in the EHR — per eight hours of patient care. Over that same period, weekly visit volume rose by 0.49 — roughly one additional patient every two weeks. And time spent in the EHR after hours didn't change at all.

Read those numbers together and a pattern emerges. The minutes saved during the day didn't shorten anyone's evening; they surfaced as slightly more throughput. The time was reclaimed — just not by the clinician, and not by the patient.

From a design standpoint, this is the gap that matters: efficiency delivered to the system is not the same as relief delivered to the clinician, or presence delivered to the patient. A tool can succeed on its own metric and still hand the saved minutes to no one who needed them.

2. Cognitive load doesn't disappear — it changes shape into something harder to see

Here's the part the well-being numbers don't capture. Even when self-reported cognitive load drops, the work doesn't vanish. It transforms — from generating a note to auditing one. And that shift is safety-relevant.

Auditing a note means catching what's missing or subtly wrong inside fluent, confident text. That's a harder detection task than writing from scratch, and it degrades precisely when the tool feels most trustworthy. This is automation complacency: the smoother the output, the more readily a tired or hurried physician clicks "Sign."

I spoke with a cardiologist who uses an AI scribe. He values being able to give his patients his full attention during a visit. But he was candid about the tradeoff: his cognitive load hasn't so much decreased as changed clothes. It's been replaced by hyper-vigilance, because he trusts his own notes more than the AI's. Writing them himself took a little longer, he said — but he felt more relaxed doing it.

His experience runs counter to the aggregate findings, and that's exactly why it's worth sitting with. He is doing the auditing the tool demands. The uncomfortable question is how many clinicians — under real time pressure, and lulled by effortless output — are doing it as carefully as he is, and what slips through when they don't.

3. When accuracy fails in a clinical scribe, the failure can be lethal

Accuracy matters in all healthcare AI, but the stakes are sharpest with the note that drives care. A scribe that records the wrong medication — or the right medication at the wrong dose — can kill someone. And the inaccuracies are not hypothetical.

A procurement study by the Auditor General of Ontario tested 20 approved AI-scribe vendors and found:

  • Nine of 20 (45%) produced hallucinations — inventing details such as therapy referrals or blood tests that were never mentioned in the simulated encounter.

  • Twelve of 20 (60%) recorded incorrect information, including capturing a different drug than the one the physician had prescribed.

  • Seventeen of 20 (85%) missed key mental-health details even when those details were clearly stated in the recording.

Every one of those failures lands back on the same requirement: a human has to catch it. Which puts the cardiologist's hyper-vigilance in a different light — not as an overreaction, but as the job.

4. The patient can't consent to what they can't see

I spoke with a psychiatrist who won't use an AI scribe even as colleagues in his practice do. His hesitation is twofold: the risk of inaccuracies we've just seen, and the risk that patient data isn't adequately protected once the conversation leaves the room. He can't independently verify what happens to it — and in his specialty, where the Ontario testers found mental-health details were the most likely to be dropped, that caution is hard to argue with.

A wave of California class actions suggests his second instinct is well-founded. In late November 2025, a patient sued Sharp HealthCare, alleging that a Sharp Rees-Stealy clinic recorded his July appointment through Abridge's ambient scribe without his consent — something he says he discovered only when he read his own visit notes afterward. California is an all-party-consent state: everyone in a confidential conversation must agree before it can be recorded. The complaint alleges Sharp had no consent process that met that bar.

The most striking allegation isn't the missing consent — it's the manufactured consent. The suit claims the AI auto-inserted statements into patient charts saying patients had been advised of the recording and had agreed to it, even when, the plaintiff says, no such conversation ever took place.

A second case, filed in federal court in April 2026, alleges much the same: three patients sued Sutter Health and MemorialCare in the Northern District of California, claiming their conversations were recorded and transmitted to third-party servers for processing without informed consent, in violation of California's Invasion of Privacy Act and several related privacy statutes. One detail is worth underlining for anyone who assumes HIPAA settles the matter: Abridge signs HIPAA business-associate agreements with its clients. The alleged problem isn't HIPAA compliance — it's that a signed BAA does nothing to satisfy a state law requiring the patient to say yes.

From a design standpoint, this is the consent dark pattern in its purest form. Consent was handled as a backend compliance artifact — a policy, a signed agreement, a box in a workflow — rather than a frontend experience the patient could see and act on. The exam room offered no visible signal that recording was underway, so the patient had no situational awareness and no genuine chance to decline. And in the Sharp allegations, the system went further still: it didn't merely fail to capture a choice, it generated a record asserting a choice the patient never made.

The thread that ties them together

One design principle runs underneath every one of these problems: these tools were built to feel effortless and to disappear. In a safety-critical setting, what disappears along with them is the signal a human needs to stay accountable — the ledger of where the saved time went, the cue that a note needs scrutiny, the indicator that the room is being recorded.

Allen Hsiao, MD, of Yale School of Medicine, frames that disappearing act as the whole point — and in the best case, he's right:

“Ambient AI is so exciting to me because it allows technology to fade into the background and allows care to come to the foreground.”

In the ideal scenario, that's exactly what happens. In the real one, the cognitive effort spent writing a note doesn't evaporate — it moves to reviewing the note. So the design task isn't to make the technology vanish completely. It's to decide what should stay visible. Designers of clinical AI scribes should build back in just enough friction — at the moments of highest stakes — to help physicians catch the hallucinations, the inaccuracies, and the quiet omissions before they reach the chart.

Keeping technology in the background is a fine goal. Keeping the human's attention in the foreground, where the stakes are highest, is the more important one.

Sources


If this resonates with you, let's talk.

Contact

© 2026 Baleen Advisory, LLC

© 2026 Baleen Advisory, LLC