AI Scribes: Why Your Doctor’s Note Is Now Written by Software, and What It Means for You

The laptop lid is down. That is the first thing many people notice at a check-up these days: the doctor is looking at them, not at a screen, and a small phone propped on the desk is quietly listening. At the end, there is a question that did not exist a few years ago. “Was the recording okay with you?”

The reason the AI scribe is trending right now is a shift from pilot to default. Through 2025, several of the largest health systems in the United States moved ambient documentation from a few hundred volunteer physicians to tens of thousands of clinicians, and in April 2025 NHS England published formal guidance on how these products should be governed. As of early 2026, a growing share of primary care and specialty visits produce a note first drafted by software.

That raises fair questions. Who hears the recording? How often does the software get it wrong? And what should you do with a note that reads differently from the conversation you remember? The evidence has answers, some of them reassuring and some of them honest about the gaps.

What is an AI scribe, and why is it suddenly in every exam room?

An AI scribe is a program that records the spoken conversation between a clinician and a patient, converts the speech to text, and then uses a large language model to turn that transcript into a structured clinical note. A large language model is a type of artificial intelligence trained on enormous amounts of text so it can summarize and rephrase language. The result lands in the electronic health record as a draft, usually within a few minutes, and the clinician edits and signs it.

The word “scribe” is borrowed from an older arrangement. For two decades, some practices hired human medical scribes, often pre-medical students, to sit in the room and type while the doctor talked. The software does the same job without a third person present, and without a lunch break.

Why now? Two things matured at the same time. Speech recognition became reliable enough to handle a coughing patient and a doctor with an accent, and language models became good enough to write passable medical prose rather than a word-for-word transcript. Health systems had a problem waiting for exactly this fix: time-and-motion studies published in the mid-2010s found that physicians spent close to two hours on electronic record tasks for every hour of face-to-face care, with a further hour or two of “pajama time” at home. Documentation burden sits near the top of nearly every physician burnout survey.

So the technology arrived into a workforce that was exhausted and a market that was ready. The result is one of the fastest adoptions of any clinical tool in recent memory, and one that most patients never explicitly chose.

What changed recently

Three dated developments explain the current wave of attention.

In January 2024, the World Health Organization published guidance on large multi-modal models in health, with more than 40 recommendations for governments and developers. It named inaccurate or fabricated output, data privacy and automation bias, meaning the human tendency to trust a machine’s suggestion too readily, as core risks. Documentation tools were squarely within scope.

In 2024, one large US health system reported results from a ten-week rollout to more than 3,400 physicians across roughly 300,000 encounters. Physicians described less after-hours charting and more attention for patients, and the report became the reference point that many other systems used to justify their own deployments. By late 2025, deployments measured in tens of thousands of clinicians were no longer unusual.

In April 2025, NHS England issued guidance on AI-enabled ambient scribing products. It set out expectations on data protection impact assessments, clinical safety processes and, notably, the point at which a product that summarizes or structures clinical information may count as a medical device and require registration. That was the first national health service to say so explicitly.

In the United States, regulators have so far treated most ambient documentation tools as outside medical device review, on the reasoning that they document rather than diagnose or recommend treatment. That line is being watched closely, because newer versions increasingly suggest diagnosis codes, draft orders or propose follow-up plans, which begins to look like clinical decision support.

Put together: a global ethics framework, a very large real-world deployment and a national governance standard all landed within about sixteen months. Search interest followed.

How does an ambient AI scribe work during a visit?

“Ambient” simply means the microphone runs in the background instead of the clinician dictating into it. Here is the sequence in a typical visit.

Before anything is recorded, the clinician opens the application, selects the appointment and starts a session. Most systems require the clinician to confirm the patient has been told. The microphone then captures everything said in the room, including small talk, the parts where you describe your week and the moment the doctor explains a lab result.

When the session ends, the audio is sent, encrypted, to the vendor’s servers. Speech recognition produces a transcript, often tagged by speaker. A language model then reads that transcript against a template, commonly the “SOAP” format: Subjective (what you reported), Objective (what was examined or measured), Assessment (the clinician’s impression) and Plan (what happens next). The model decides what belongs in each part, drops the chatter and rewrites your words into clinical language. “My knee has been killing me since I slipped on the porch” becomes “Patient reports left knee pain since a fall two weeks ago.”

The draft appears in the clinician’s record system, frequently within two to five minutes. The clinician is expected to read it, fix errors, add anything the exam revealed that was not spoken aloud and sign. Only a signed note becomes part of your legal medical record.

What happens to the audio afterwards varies by vendor and contract. Some delete the recording within hours of the note being signed; others keep audio and transcript for a defined period for quality checks. That retention policy is the single most useful thing to ask about if privacy is your concern.

Why are doctors switching to AI scribes so quickly?

Ask a physician why they adopted an AI scribe and the answer is rarely about technology. It is about Sunday night.

Documentation does not fit inside a fifteen-minute appointment. Notes get finished between patients, over lunch and, most often, at home after dinner. Surveys of US physicians across specialties have repeatedly linked this unfinished charting to burnout, and burnout to intentions to cut hours or leave medicine. The economics are blunt: replacing a departing physician is far more expensive for a health system than any software license, and the system cannot recruit its way out of a national shortage.

Early users describe three concrete changes. First, the keyboard goes quiet during the visit. Clinicians report making eye contact, noticing body language and asking a follow-up question they would once have skipped while typing. Second, the note is drafted while they walk to the next room, so the end-of-day backlog shrinks. Third, the writing itself improves in a specific way: human-written notes under time pressure tend to be terse or copied forward from the last visit, while an AI draft captures what was actually said this time.

There is a less flattering reason too. Documentation also drives billing and quality reporting, and a fuller note can support a higher billing code. Health systems rarely lead with this, but it is part of the calculation.

None of these benefits are automatic. Clinicians who skim the draft and sign without reading lose the safety check that the whole arrangement depends on. The best deployments pair the software with training on how to edit, not just how to press record.

What the evidence actually says

The honest summary is that the evidence for AI scribes is promising, mostly observational and still young. It helps to grade it.

Strongest signal, moderate quality: clinician experience. Multiple pre-and-post surveys from large deployments, several published in peer-reviewed journals during 2024 and 2025, show meaningful drops in self-reported documentation burden and burnout scores among regular users, along with reductions in after-hours record time measured from system logs. These are real-world data, which is a strength, but they lack randomized control groups, and clinicians who volunteer for new tools tend to be enthusiasts. Early randomized trials comparing clinicians assigned to use a scribe with those who were not began appearing in 2025; they are small and short, and their results so far point the same way as the observational data, which is encouraging but not conclusive.

Moderate signal, mixed quality: time saved. Studies measure different things, so figures range from a few minutes per note to an hour or more per day. Clinicians who type slowly or write long notes gain most; fast typists sometimes gain little.

Weak signal: patient outcomes. There is no published randomized evidence that patients whose notes were drafted by an AI scribe have better health outcomes, fewer errors in care or higher satisfaction, although patient surveys from pilots are generally positive about attention and eye contact.

Open question: safety after signing. Audits of generated notes consistently find omissions and occasional fabrications, discussed in the next section, but nobody has yet shown whether those errors propagate into wrong prescriptions or missed follow-up more often than errors in human-written notes.

Expert opinion, including WHO’s 2024 guidance, favors adoption with human review, measurement and transparency. That is roughly where the field stands.

AI scribe accuracy: how often do the notes get things wrong?

Every published evaluation of AI scribe output finds errors. The useful questions are what kind, how often and who catches them.

Researchers typically compare generated notes against the audio transcript and sort problems into three bins. Omissions are the most common: the patient mentions a new over-the-counter supplement, a recent fall or that a symptom started on the right side, and the note leaves it out. Distortions come next: the duration of a symptom shifts from “a few days” to “a few weeks”, or a tentative impression is written as a firm diagnosis. Fabrications, often called hallucinations, meaning confident statements with no basis in the conversation, are the rarest category but the one safety researchers worry about most, because they are hard for a tired clinician to spot. A classic example is a normal physical exam finding being documented when no exam of that body part took place.

Rates differ by product, specialty, background noise and how many people are talking. Pediatric visits with a parent, child and sibling in the room, or visits conducted through an interpreter, are harder. Medication names are a known weak point for speech recognition because so many sound alike.

Two things keep this in perspective. Human-written notes also contain errors; studies of open notes have found that about one patient in five who reads their record spots a mistake, and roughly 40 percent of those call it serious. And the scribe is a draft, not a record. The designed safeguard is the clinician reading before signing.

The practical consequence for you is simple. The note is more likely than before to reflect what you said, and slightly more likely to contain a confident sentence that nobody said. Reading it is worth ten minutes.

Is an AI medical scribe HIPAA compliant?

HIPAA, the Health Insurance Portability and Accountability Act, is the US federal law that sets rules for how health information is protected and shared. A frequent search is whether an AI medical scribe is “HIPAA compliant”, and the precise answer is that HIPAA does not certify products; it regulates the people and organizations that handle your data.

In practice, a clinic or hospital that uses an AI scribe must sign a Business Associate Agreement with the vendor. That is a legally binding contract requiring the vendor to protect your information to HIPAA standards, report breaches and use the data only for the agreed purpose. Reputable vendors sign these routinely, encrypt audio in transit and at rest, and restrict which employees can access recordings. Many also commit, in writing, not to train their models on identifiable patient data without separate consent. That last point is worth asking about, because “HIPAA compliant” in a brochure does not by itself tell you whether your voice is being used to improve the product.

Recording consent is a separate layer of law. Federal rules allow recording when one party to the conversation consents, which the clinician does. Roughly a dozen states, however, require all parties to agree before a conversation is recorded, which is why health systems in those states ask you explicitly. Many systems ask everywhere as a matter of courtesy and policy.

Outside the United States, the picture is stricter. NHS England’s 2025 guidance requires a data protection impact assessment before deployment and expects products that structure clinical content to meet medical device standards.

So the fair answer is: a well-governed deployment can meet HIPAA’s requirements, the paperwork exists to prove it, and you are entitled to ask to see the privacy notice that describes it.

Will you be asked for consent, and can you say no?

You should be told, and in most settings you will be asked. The form that takes varies: a sign at reception, a line in the intake paperwork, a verbal question at the start of the visit, or all three. Good practice, reflected in both WHO’s 2024 guidance and NHS England’s 2025 standard, is that the explanation happens before the microphone is switched on and that saying no carries no penalty.

Declining is your right, and clinicians are instructed to accept it without argument. The visit then proceeds the way it did before: the doctor types, dictates afterwards or writes from memory. Some patients decline for everything; many decline for specific conversations, such as a mental health visit, a sexual health concern or a discussion involving a family member, and accept for a routine blood pressure check. That selective approach is entirely reasonable and worth knowing about, because consent is per visit, not forever.

Four questions will tell you most of what matters, and a clinician using the tool properly should be able to answer them in a sentence each:

• Who is the company that processes the recording, and is there a signed agreement protecting my data?
• How long is the audio kept after the note is signed?
• Is my recording used to train or improve the software?
• Will the doctor read the note before signing it?

If the answers are vague, that is information too. It does not mean the tool is unsafe, but it suggests the practice has not thought carefully about explaining it, and you can reasonably ask for the written privacy notice.

One more point for anyone who brings a companion: the recording captures their words as well, so they should be asked too.

What an AI scribe changes about your appointment, and what it does not

The most visible change is attention. Patients in pilot surveys describe the doctor as more present, and clinicians say they ask better follow-up questions when they are not typing. If you have spent years talking to the side of someone’s head, that alone feels like a different kind of medicine.

A subtler change is in the written record. Because the scribe captures the conversation rather than the clinician’s compressed memory of it, your own account tends to survive in fuller form. The sentence “I worry this is what my mother had” is more likely to appear in some version, and that context can shape care at the next visit, perhaps with a different doctor.

Some things do not change at all. The visit is not longer; health systems have generally kept appointment slots the same length and used the saved time to shrink backlogs rather than extend consultations. The software does not examine you, does not order tests and does not decide anything. The clinician’s judgement is still the product; the scribe is stationery.

And a few things get slightly harder. Speaking over one another confuses speaker tagging, so you may notice the doctor letting you finish, which is no bad habit. Jokes and idioms occasionally land in the note in odd ways. Patients who switch languages mid-sentence, or who speak quietly, are documented less accurately, and the published evidence on performance across accents and languages is thin, which is a genuine equity concern.

If you want the recording to work well for you, the practical advice is to speak as you would anyway but to name things clearly: say the supplement’s name, the side of the body, and how long. The software, like the doctor, documents what it hears.

AI scribe vs human scribe vs dictation: how the options compare

Doctors have had several ways to get words into the record, and each trades something for something else. The table summarizes what the evidence and everyday experience suggest.

Method	Who writes the first draft	Typical strengths	Typical weaknesses	Privacy footprint
Clinician types during visit	Clinician	No third party; immediate control	Divided attention; terse or copied-forward notes; after-hours work	Record system only
Dictation after visit	Clinician speaks, software transcribes	Clinician chooses every word	Relies on memory; still done after hours	Transcription vendor
Human scribe in room	Trained person	Can ask the clinician to clarify in real time; captures exam	Extra person present; staff turnover; variable skill	Scribe plus record system
Ambient AI scribe	Language model from recording	Fast draft; full conversation captured; keyboard closed	Omissions and occasional fabrications; cannot see the exam; needs careful review	Audio and transcript held by vendor for a defined period

Two patterns stand out. The human scribe is the only method that can see what the doctor does, not just what is said; an AI scribe cannot document a swollen ankle unless someone describes it aloud, which is why clinicians using the tool often narrate their exam. And the ambient method is the only one that creates a recording of your voice that leaves the building, which is why its consent and retention rules deserve more scrutiny than the others.

Comparative studies pitting these methods directly against each other on note accuracy are few. What exists suggests AI drafts are more complete than hurried typed notes and roughly comparable to good human scribes on content, while introducing a new error type that human scribes rarely make: the confident invented sentence.

Which AI scribe is best, and is an AI scribe free?

These two questions dominate search results, mostly from clinicians shopping for a tool, and the honest answers are shorter than the sales pages suggest.

There is no independent, head-to-head benchmark that establishes a “best” AI scribe. Vendors publish their own accuracy figures using their own test sets, and academic evaluations have generally studied one product at a time inside one health system. A tool that performs well in a quiet dermatology clinic may struggle in a noisy pediatric urgent care. For a patient, the relevant question is not which brand the practice chose but whether the practice has done three things: signed a data protection agreement, trained its clinicians to review drafts and told you clearly how recordings are handled. Those are governance questions, and good governance matters more than the logo.

On cost, the answer for patients is unambiguous: you are not charged for it, and it should never appear on a bill. Whether the practice pays, and how much, depends on contracts that vary enormously, and some vendors offer limited no-cost tiers aimed at individual clinicians. That last detail deserves a caution rather than a recommendation. Tools adopted informally by an individual doctor, outside the organization’s agreements, may lack the signed data protection contract described earlier. If your doctor is using a scribe on a personal device, it is fair to ask whether the practice has approved it.

The question clinicians should really be asking is not “which is best” but “how will we measure it after launch”. Error rates, edit time, patient complaints and whether notes are actually being read before signing are the numbers that tell a practice whether its scribe is working. Few deployments publish them yet. The ones that do are the ones to trust.

Common myths about AI scribes, corrected

Viral posts about AI scribes tend to swing between miracle and menace. Neither survives contact with the published evidence.

Myth: “The AI is writing my diagnosis.” The software summarizes a conversation; the clinician forms the diagnosis and is legally responsible for the signed note. Where products suggest billing codes or plans, those suggestions are still accepted or rejected by a human, and regulators in both the US and England are watching that boundary.

Myth: “My recording is kept forever and sold.” Business Associate Agreements and, in Europe, data protection law restrict what vendors may do. Retention periods exist and are disclosed on request. Selling identifiable recordings would be a reportable breach. What is true is that policies on using de-identified data for product improvement vary, and you can ask.

Myth: “AI notes are more accurate than doctors’ notes.” Audits show they are more complete than rushed typed notes and also contain a new type of error. “Different” is the defensible word, not “better”.

Myth: “If I decline, my care will suffer.” There is no evidence of that, and declining is explicitly protected in professional guidance. Many clinicians treated patients for decades without the tool.

Myth: “The doctor doesn’t read the note anymore.” Some do not read carefully, and that is the real risk. But the design intent, every vendor’s instruction and every piece of guidance say the same thing: read before signing. Pointing out an error in your portal is one of the strongest nudges toward that habit.

Myth: “It has proven to reduce burnout.” It has shown reduced burnout scores in observational studies and early trials. That is a genuine signal, not proof, and the difference matters when deciding how much to invest in it.

How to read and correct your AI-written visit note

Since 2021, US federal rules under the 21st Century Cures Act have required most clinical notes to be shared with patients promptly, usually through a patient portal, an online account where you can view your records. That means you will often see the AI-drafted, clinician-signed note within a day of your visit. Here is how to read it well.

Start with the Plan section, because that is where errors do harm. Check that the medication names, the tests ordered and the follow-up timing match what you discussed. Then read the Subjective section, your own account, and look for the three error types: something you said that is missing, something that has shifted (left became right, days became weeks), and something you never said at all. Finally glance at the Objective section for exam findings that could not have happened, such as a normal abdominal exam when you kept your coat on.

If you find a problem, you have a formal right under HIPAA to request an amendment to your record. The provider must respond within 60 days, and if they disagree they must tell you why and let you add a statement of disagreement to the file. In practice, a short portal message to the clinic describing the error is usually enough, and most practices will correct a clear factual mistake without the formal process. Keep it specific: “The note says I take a daily aspirin. I stopped that in March.”

Do not rely on the note as medical advice, and do not change anything you take or do because of a sentence in it. If the note and your memory disagree about an instruction, that is a reason to ask, not to guess. The clinician who signed it is the person who can tell you which version is right.

When to see a doctor: red flags in an AI-written note

Most note errors are harmless, a misspelled town or a cat described as a dog. A small number are not, and they deserve a prompt call rather than a shrug.

Contact the practice the same day if your note contains any of the following:

• A medication listed that you do not take, or one you do take that is missing, since other clinicians and pharmacists read this list when prescribing.
• An allergy or adverse reaction recorded that you have never had, or a known allergy that has disappeared.
• The wrong side of the body, the wrong limb or the wrong organ in a note that will guide a procedure or imaging.
• A diagnosis you were never told about, or a condition written as confirmed that the doctor described as a possibility.
• A follow-up instruction, test or referral that differs from what you agreed, or that is absent entirely.
• A symptom you reported, such as chest pain, shortness of breath, bleeding, a new lump or sudden weakness, that does not appear in the note at all, because missing documentation can mean a missed workup.

Some situations should not wait for a note correction. If you develop new or worsening chest pain, difficulty breathing, signs of a stroke such as facial drooping, arm weakness or slurred speech, a severe allergic reaction, or heavy bleeding, seek emergency care immediately regardless of what any record says.

And a boundary worth stating plainly: never start, stop or adjust a medicine because of what a note says or fails to say. The decision about your treatment belongs to the clinician who knows you, and a documentation error is a reason to ask them, in person or through the portal, not a reason to act alone. The scribe wrote the draft; your doctor is still the one who answers for it, and the one who can put it right.

Frequently asked questions

What is an AI scribe in a doctor's office?

An AI scribe is software that records the conversation during a medical visit, turns the speech into text and uses a language model to draft a clinical note. The clinician reviews and signs the draft before it enters your record. It replaces typing during the visit, and some practices use it instead of a human scribe who previously sat in the room.

Is an AI scribe free for patients?

Yes, in the sense that patients are never charged for it and it should not appear on a bill. The practice or health system covers the technology through its own contracts, which vary widely. If you ever see a line item that looks related to recording or documentation software, ask the billing office, because that would be unusual and worth questioning.

Which AI scribe is best?

No independent head-to-head benchmark exists, so there is no evidence-based answer. Vendors publish their own accuracy figures, and academic studies usually evaluate one product in one setting. For patients, what matters more than the brand is whether the practice has a signed data protection agreement, trains clinicians to review drafts and explains clearly how recordings are stored.

Is an AI medical scribe HIPAA compliant?

HIPAA does not certify products; it regulates the organizations handling your data. A practice using an AI scribe must sign a Business Associate Agreement obliging the vendor to protect information to HIPAA standards and report breaches. Well-governed deployments meet those requirements, and you are entitled to ask for the privacy notice that describes them.

Can I refuse the AI scribe recording?

Yes. Professional guidance in the US and England says patients must be told before recording starts and may decline without any effect on their care. You can decline for a single sensitive visit and accept at the next one. If you refuse, the clinician simply documents the way they did before, by typing or dictating afterwards.

How accurate is an ambient AI scribe compared with a human?

Audits suggest AI drafts are more complete than hurried typed notes and broadly comparable to good human scribes on content, but they introduce a new error type: confident sentences nobody said. Human scribes can also see the physical exam, which the software cannot. Accuracy drops with background noise, multiple speakers and interpreter-mediated visits.

Does the AI scribe diagnose me or decide my treatment?

No. It summarizes what was said; the clinician forms the diagnosis and chooses the plan, and is legally responsible for the signed note. Some newer products suggest billing codes or follow-up items, but a clinician accepts or rejects those. Regulators are watching that boundary because suggesting care begins to look like clinical decision support.

What happens to the recording of my visit afterwards?

It depends on the vendor and the contract. Some delete audio within hours of the note being signed; others retain audio or transcript for a defined period for quality checks. Policies on using de-identified data to improve the software also vary. Asking how long audio is kept and whether it trains the model will tell you most of what you need.

What is AI scribe accuracy like for children's visits or interpreters?

Lower than for a quiet one-to-one adult visit, according to the limited data available. Pediatric encounters often involve a parent, a child and siblings talking at once, and interpreter-mediated visits add a second language, both of which challenge speaker tagging. Published evidence on performance across accents and languages remains thin, which researchers flag as an equity gap.

What should I do if my AI-written note contains a mistake?

Send a short, specific message through the patient portal or call the clinic, especially for errors in medications, allergies, body side or follow-up plans. You have a HIPAA right to request an amendment, and the provider must respond within 60 days. Do not change any medicine based on the note; ask the clinician which version is correct.

References

• WHO: Ethics and governance of artificial intelligence for health: guidance on large multi-modal models
• WHO: Ethics and governance of artificial intelligence for health
• MedlinePlus: Personal Health Records

This article is for general information only and is not a substitute for professional medical advice. Please consult a qualified doctor about your individual situation.