My clinic's portal sent me a notice last year — one paragraph, easy to skip — saying my care team might use an AI assistant to help with documentation during visits. I read it. Something small went off in me, the way it does. Then I dismissed it in about four seconds, on reasoning I can reconstruct exactly: they already hold my entire medical history, a microphone is not the frightening part of that, and my doctor looks exhausted and I want her to keep doing this job.
Every link in that chain was correct. The conclusion was still wrong.
Because what I was weighing was the recording, and the recording is not what follows you. What follows you is a paragraph of prose about your body — drafted by a language model, signed by a human running eleven minutes behind, and filed where every clinician you meet for the rest of your life will read it before they meet you. I had never read one of mine. I would guess you have never read one of yours.
The thirteen minutes are real. They are also not the thirteen minutes you were sold.
Fairness to the technology first: this is the rare case where the AI roughly did what it promised.
On April 1, 2026, a study of 8,581 clinicians published in JAMA reported that adopting an ambient AI scribe — software that listens to the visit and drafts the clinical note — was associated with 13.4 fewer minutes of electronic health record time, 16.0 fewer minutes of documentation time, and about half an extra patient visit a week. It tracked five academic health systems over two years, came from a collaborative of institutions rather than a vendor, and had Lisa Rotenstein of UCSF as first author.
Now the fine print.
Those thirteen minutes are per eight scheduled patient hours. Not per patient. Not per visit. Per roughly a full clinic day. Read as a saving on every appointment, the number sounds transformative. It is not. It is a clinic day that ends thirteen minutes earlier, which Mass General Brigham, one of the five sites, called "modest daily reductions." The researchers' own word, used twice in their conclusion, is modest.
Three more caveats sit inside the same paper. It is an observational difference-in-differences cohort covering adoptions between June 2023 and August 2025 — clinicians chose whether to adopt, so this supports association, not causation. Electronic health record time outside work hours did not change significantly: the evenings-and-weekends charting doctors call pajama time stayed exactly where it was. And the gains were greatest for clinicians who used AI scribes in 50% or more of their visits. STAT called the savings modest and the usage inconsistent — that second word is STAT's headline framing rather than a finding the paper spells out, though the paper does substantiate this much: it works if you use it.
So far, so good. And it is already everywhere: as of June 2025, nearly two-thirds of the hospitals running Epic had adopted an ambient documentation tool.
Here is where it gets interesting.
What is actually in the paragraph
The research that produced the thirteen minutes also produced numbers about what the machine writes, and they disagree by an order of magnitude. So let me attach each to its method.
Roughly 1% to 3%. A commentary in npj Digital Medicine reports that scribes built on large language models show lower overall error rates than older systems, while introducing new failure modes — hallucinations, critical omissions, misattribution, contextual misinterpretation. Its authors warn that adoption is outpacing validation and oversight, and that the rush to deploy "may compromise patient safety, clinical integrity, and provider autonomy." Two details matter later: these systems can log a patient's statement as the clinician's, and the underlying speech recognition shows higher error rates for African American speakers.
11.5% and 18%. A UC Davis pilot that had physicians grade 356 AI-drafted notes found accidental omissions in 18% of them and hallucinations in 11.5%. Its authors are balanced: 94.7% of notes were free of significant errors, but 5.3% carried errors rated as posing serious or imminent risk, and careful clinician review "remains imperative." Which is why a different number there is the one I cannot stop thinking about: 14.9% of notes were signed with no edits at all — one in seven going into the permanent record exactly as the model wrote it.
31% versus 20%. In a comparison run by researchers affiliated with the scribe vendor Suki AI, hallucinations turned up in 31% of ambient notes against 20% of physician-written "gold" notes — striking for a vendor-affiliated team to publish, and also one study of 97 encounters whose authors call their sample size "modest," whose comparison notes came from single physicians working off the same audio, and whose reviewers still preferred the ambient notes (47% to 39%). Not the hallucination rate. One data point, from an interested party, pointing where the others point.
And on quality, humans still win. A Veterans Health Administration and University of Washington team scored eleven AI scribes against eighteen human clinicians in Annals of Internal Medicine; the human notes scored higher in every case. And a randomized trial at UCLA that found one commercial scribe cutting documentation time by nearly 10% also reported physicians judging the notes to "occasionally" contain clinically significant inaccuracies, most commonly omissions, with one mild patient safety event.
So: the tool works, and what it produces is a draft. The safety model rests entirely on a human catching what the model got wrong — and in the one study that counted, one note in seven was never edited.
And a second question: did you agree to this?
In December 2025, a proposed class action alleged that a San Diego health system, Sharp HealthCare, recorded doctor-patient conversations using an AI tool and falsely documented that patients had consented; it estimates more than 100,000 patients may have been recorded. Plaintiffs claim it amounts to electronic eavesdropping under California's wiretapping statute, even though no human ever listened. These are allegations: nothing has been found or proven, and both sides said they cannot comment on pending litigation.
But one detail stopped me cold. The named plaintiff, Jose Saucedo, says he discovered the alleged recording only after reading his own medical notes from a July appointment. Whatever a court decides, the mechanism by which a patient found out was not a disclosure. It was him, reading his chart.
So that settles it — these things are dangerous. Right?
No — and I have to argue against my own framing, because the case for ambient scribes is stronger than the error rates make it sound.
Across six health systems, burnout fell from 52% to 39% among 263 clinicians after thirty days with an ambient scribe. At Mass General Brigham and Emory, more than 1,400 clinicians using ambient documentation reported significantly improved well-being, with burnout dropping from 51% to 29%.
I will not launder those into nothing — but the authors' own caveats travel with them. The six-system study had no control group to adjust for temporal trends, drew on early adopters, and ran a survey that was not anonymous; participants, the authors note, "may have responded favorably to please their digital health leadership." Convenience samples of enthusiasts at thirty days are the softest evidence in medicine, and nearly the only evidence we have. Burnout is not a soft problem either: exhausted physicians make worse decisions and leave. The Peterson Health Technology Institute's summary of what health system leaders report is the most honest sentence about this technology: the tools likely improve clinician burnout, but the financial impact is unclear.
So I am not arguing that anyone should take these tools away. A doctor who is not typing is a doctor who is looking at you. My argument is narrower and more annoying: if the safety of the arrangement depends on somebody reading the paragraph, somebody had better read it — and the second reader, the one with the most at stake, is you.
Sweden has the same machine in the room. What differs is who reads the paragraph.
I keep coming back to Sweden, and not because Swedes are tidier than the rest of us. They are running the same experiment — they just started the other half of it fourteen years ago.
The machine is already there. Karolinska Institutet reported in June 2026 that during tens of thousands of appointments across the country, an AI sits in as a silent third party, transcribing before producing a record entry a physician signs; the recording is deleted afterwards. One vendor, Tandem Health, has sold its system into all of the Gävleborg region's primary care.
What differs is the other end of the pipe. Swedes have read their record notes online since Region Uppsala gave all its patients access in November 2012, and every region was connected by 2018. Inera, which runs the service, states the rule simply: every resident aged 16 and over can read their notes. You log in with a national e-ID — though how much you see still varies by region and clinic.
Underneath the app sits a statute. Sweden's Patient Data Act of 2008 provides that if a patient considers information in the record incorrect or misleading, it shall be noted in the record, and that a record shall on request be made available to read as soon as possible. Sweden's health and social care inspectorate restates it plainly: you almost always have the right to read your record, and if the provider decides an entry is correct and refuses to change it, you may demand a notation that you consider it wrong. 1177, Sweden's public health service, adds the honest limit — you do not get to write in your record or decide what it says. You get to be on the page.
What do they find? In a nationwide survey of Swedish mental health care patients who read their records, half of them found an error and about a third found an omission — self-reported, with the response bias the authors flag. And Swedes do read: in a four-country Nordic survey of 29,334 portal users, Swedes read theirs more than anyone, a third logging in over twenty times. Maria Hägglund of Uppsala University sums up the decade as patients gaining "a better understanding of their health."
Then the regulatory mirror image. Sweden's Medical Products Agency announced on March 4, 2026 that during spring 2026 it plans to carry out inspections of the manufacturers of these systems, looking at how medical content is generated and quality-assured. Its position is that a system which does not merely transcribe but compiles patient-specific medical information later used as a basis for care decisions is probably covered by medical device rules — "troligen," in the original. In June the agency was given responsibility for market surveillance of high-risk AI under the EU's AI Regulation.
Two limits, before anyone puts Sweden on a pedestal. I found no evidence it passed any rule specific to AI-generated notes — the contrast is that the general right to read applies to whatever ends up in the record, including the machine's draft once a human signs it. And Sweden is not tidy: a survey of 357 Swedish physicians found that Sweden has its own shadow problem, clinicians using unauthorized AI tools that never passed conformity assessment.
Same machine. Same mess. A different default about who gets to look.
Now run this forward five years
It is 2031, and your notes are no longer documents that get filed. They are inputs. A coding engine reads them to bill. A prior-authorization model decides whether your insurer pays for the scan. A triage model reads them before the nurse does.
In that world an omission stops being a blank space and becomes evidence. If the note never mentions the numbness in your left hand, the model downstream does not read "unknown" — it reads "not reported," and treats absence as a negative finding. Multiply that across five years of visits and you get a medical history with holes shaped like the things the model did not think mattered.
Now run misattribution forward. In the last ninety seconds of an appointment you mention that you have been drinking more since your father died. The note records increased alcohol use. Nobody asks you about it again — not because anyone is careless, but because the answer is already in the file, and the file is what people read when they are running behind.
The optimistic 2031 is just as plausible, and it turns on something embarrassingly small: patients reading their notes as routine, the way Swedes already do. The difference between those two futures is not a breakthrough. It is a habit.
What the people who study this are converging on
Read across the ideological spectrum and the striking thing is how little disagreement there is — and how little is happening in Washington.
In January 2026, the FDA pulled back on clinical decision support software: products delivering a single recommendation, once treated as regulated medical devices, can now reach the market without FDA review if they meet the agency's other criteria. Arnold & Porter reads the 2026 CDS Final Guidance as broadening what counts as non-device software. Meanwhile states, not Washington, are writing the rules on health care AI, which is how you end up with a patchwork.
Against that backdrop, look at who is asking for the same thing. From bioethics, The Hastings Center argues that patients may not receive adequate information about ambient AI before it is used — where the data is stored, whether their speech is analyzed — limiting their ability to consent meaningfully. From the market-oriented side, the Paragon Health Institute, no friend of heavy premarket regulation, wants post-deployment scrutiny added through a public-private partnership. When even the deregulators want post-deployment monitoring, the question has stopped being partisan. And the doctors: the American Medical Association's own 2026 survey of 1,692 physicians — over 80% of whom now use AI professionally, double the 2023 share — found they rank clear liability frameworks for errors and post-market surveillance as their top regulatory priorities. Not a ban, not a pause — a way to find out whether the thing works after it is installed.
Which is, more or less, what Sweden's drug regulator says it plans to start doing this year.
What does this mean for you?
Practical, and mostly free. The strongest safeguard here is a patient who reads.
Read your last visit note this week. Not the summary card — the narrative note the clinician signed. It is in your portal, because federal rules against information blocking are the rule that made your notes yours to read.
Expect to find something. Of 22,889 US patients who read their ambulatory notes in one study, one in five reported a perceived mistake and 42.3% of those called it serious, diagnosis errors being the most common serious kind. That research predates ambient AI, and its authors are careful that these are patient-perceived errors, not independently verified ones. Notes have always had mistakes; patients have always found them.
Check the load-bearing lines first. Medications and doses. Allergies. Diagnoses, current and past. Family history. Anything about alcohol, drugs, or mental health — the lines that quietly change how the next stranger reads you.
Ask whether an AI drafted it, and what happens to the audio. At the desk, in the room, or through the portal. You are allowed to ask, and the answer tells you something either way.
If something is wrong, use the mechanism that already exists. Under 45 CFR 164.526 you can ask your provider to amend protected health information in your record, and they must act within 60 days (one 30-day extension allowed). They can deny it if they determine the information is accurate and complete — but must then let you file a written statement of disagreement that travels with the record. Weaker than Sweden's rule, where your objection shall be entered. Not nothing, and almost nobody uses it.
Do not make this a fight with your doctor. She did not buy the software or set the fifteen-minute slot, and her burnout numbers are the best argument the technology has. You are the second reader, not opposing counsel.
The lesson, as I see it
The fear I dismissed in four seconds was real. I had simply aimed it at the wrong object.
Everybody is arguing about the microphone — in courtrooms, consent forms, the alarmed group chats of careful friends. But the microphone is transient and the paragraph is permanent. The paragraph is what gets read by the specialist you see in 2029, by the emergency physician who meets you unconscious, by the model deciding your prior authorization. And it is now, increasingly, a first draft written by something that cannot be embarrassed, cannot be sued, and does not know when you were joking.
I am not against the thirteen minutes. I want my doctor to have them. But those minutes were bought with a document, and the deal only holds if somebody reads it — and the current plan is that a tired clinician catches every error at the end of a long day.
Sweden's answer was not clever and it was not technological: a law saying your objection goes in the file, and a website where you can log in and look. You already have most of that. You just have to open it.
Forward this to someone who has a patient portal login and has never once opened it — that is the entire call to action. The HAIA Foundation works to make the documents that quietly decide things about us readable, and this newsletter is where that work gets written down.