AI Scribes for Doctors: Six Note Errors to Catch Before You Sign

AI scribes make six kinds of error in doctors' notes: omissions, invented findings, wrong side or dose, negation flips, speaker mix-ups and template carry-over. Catch them by reading the plan and medicines first, then every negative and every side, then the history, comparing each against what you remember, before you sign.

The dangerous errors are rarely typos. They are fluent, confident sentences that read like your own writing, which is why a skim feels safe and isn't. A scribe produces probable prose from speech; probable is not the same as true, and the gap sits in exactly the details that matter most in a clinical record.

Follow me on Instagram@sagnikteaches

A consultation with six errors planted in it

Everything below hangs on one illustrative consultation, so you can see each error in context. Say a solo general practitioner sees a 58-year-old man for 22 minutes. He has had left knee pain for three weeks, takes ramipril 5 mg daily, has no chest pain or breathlessness, and mentions his wife has diabetes. The doctor says she will increase his ramipril to 10 mg, arrange an X-ray of the left knee and review him in two weeks. The scribe is run with a standard consultation template.

Connect on LinkedInSagnik Bhattacharya

The draft comes back in under a minute, tidy and well structured. Here are the key lines, with the real conversation beside each.

Subscribe on YouTube@sagnikteaches

That is seven problems in a note of about 180 words, drawn from the six categories (wrong side and wrong dose share one). No single line looks odd. A doctor reading at normal speed would likely sign it. The rest of this tutorial takes each error type in turn: how it happens, how to spot it fast, and what to change so it happens less.

1. Omissions: what you said that never reached the page

Scribes compress. A 22-minute conversation becomes a note, so something is dropped, and the model decides what counts as clinically relevant. It tends to keep what sounds like a complaint, a finding or a decision, and drop advice, safety-netting, conditions attached to a plan and anything said while the patient was dressing or the doctor was typing.

In the example, the bloods booking was a conditional follow-up: "I'll book kidney bloods in a week because of the dose change". It was said in passing, with no obvious noun like a drug or a body part, so it fell out. If the patient later has a raised creatinine nobody checked, the record says nothing was planned.

How to catch it

Do not read the draft looking for wrong things. Ask instead a short list of questions, and look for the answers on the page:

• Did I tell the patient when to come back, and is that date or interval written down?
• Did I give red-flag advice ("come back sooner if...")? Is it in the plan?
• Did I order or promise a test, referral or prescription? Is each one listed?
• Did the patient raise a second problem near the end? Is it there?

The last one matters most. The "by the way, doctor" complaint at the door is the classic casualty of summarisation, and it is often the real reason the patient came.

How to reduce it

Say your plan aloud in a closing sentence, slowly and in full: "So to summarise: increase ramipril to 10 mg, kidney bloods in a week, left knee X-ray, review in two weeks, and come back sooner if the knee locks or swells." Scribes weight the end of a consultation heavily, and a spoken summary gives the model every item in one place. It also helps the patient, which is the real reason to do it.

2. Invented findings: examination you did not perform

This is the error that frightens clinicians most, and the one with the most obvious consequence in a dispute. A scribe sees a note template with an "Examination" heading and a conversation that mentions a knee. It writes what an examination of a knee usually contains. "Full range of movement" is a sentence the model has seen thousands of times, and it fits.

You will see it most often in three places: examination sections, observations such as blood pressure or temperature you did not state aloud, and diagnoses you were considering but had not made ("likely osteoarthritis" when you said "could be arthritis, could be a meniscal problem").

How to catch it

Treat every finding as guilty until you can point to a moment you remember. A fast rule: for each sentence under Examination, ask "did I do that, and did I say the result out loud?" If you did it but did not speak the result, the scribe could not have heard it. A result that appears without having been spoken was invented, or inferred from the rest of the conversation.

Numbers deserve special suspicion. A blood pressure of 128/82 in a note when you only said "pressure's fine" is a made-up reading in a legal record.

How to reduce it

Dictate findings as you go: "Left knee, mild effusion, tender medial joint line, flexion limited to about 100 degrees." This is the one habit with the strongest effect. It gives the model real material, so it has less reason to supply plausible filler. Choose a template whose examination field stays empty if nothing was said, rather than one that prompts for typical findings. If your tool can show which part of the transcript a sentence came from, use that view whenever a finding looks unfamiliar.

3. Wrong side and wrong dose: small tokens, large consequences

Left and right, 10 and 15, milligrams and micrograms, once daily and twice daily. Speech recognition hears these as sounds, and the model then tidies them. "Left" and "right" are rarely confused by the ear but are easily flipped by the writer, particularly when the conversation has used both sides (you might mention the right knee as the comparison for the left, so both words appear).

Units and frequencies fail in a different way. "Take one twice a day" can become "BD" or "twice daily" correctly, but "ten of the tablets" can become "10 mg" when the pack is 5 mg tablets and she meant two. The scribe has no pack in front of it.

How to catch it

Make this a separate, deliberate pass of its own, not part of a general read. Go straight to every number and every laterality word and check each against your memory and, for medicines, against the prescribing screen. Look at the drug name, the strength, the route, the frequency and the duration, in that order. Do not trust a line because it matches the rest of the sentence.

A trick that works: read the numbers aloud, or read the note from the bottom up. Reversing the order breaks the fluent reading rhythm that lets wrong tokens slide past.

How to reduce it

Say laterality and strength in the same phrase every time, in full: "left knee", "ramipril ten milligrams once daily". If a prescribing system sits beside your notes, treat the prescription as the source of truth and have the note state the dose from it. Never let an ambient note be the only place a dose is recorded before a prescription goes out.

4. Negation flips: "no chest pain" becomes "chest pain"

Negatives are fragile. A short "no" or "denies" is easy for a transcription step to lose in a noisy room, and a summariser that has seen "chest pain" and "exertion" in the same passage may combine them into a symptom. In the example, the line "Patient reports chest pain on exertion" sits inside a note that otherwise reads calmly. Someone triaging from the record would be sent in the wrong direction.

The reverse also happens: "no history of diabetes" reduced to "history of diabetes" because the negating word sat in a previous sentence, or "not allergic to penicillin" written as an allergy. Pertinent negatives are the sentences a defence relies on later, because they show you asked.

How to catch it

Scan for the symptoms that would change management if present, and read the whole sentence each time. Red-flag symptoms are the ones to hunt: chest pain, breathlessness, weight loss, blood in stool, new neurological signs, suicidal thoughts, allergies. For each, ask "is this stated as present or absent, and does it match what I heard?"

It helps to use a template with a dedicated "Pertinent negatives" or "Denies" line. A flip then shows up as a symptom sitting in the wrong list, which is much easier to see than a missing word inside a paragraph.

How to reduce it

Speak negatives clearly and give them their own sentence. "No chest pain. No breathlessness. No ankle swelling." survives noise far better than "no chest pain or breathlessness or ankle swelling". Microphone position helps more than people expect: a lapel or desk microphone close to the patient catches a mumbled "no" that a phone across the room does not.

5. Speaker mix-ups: the wife's diabetes in his record

When two people talk, the scribe must decide who each sentence belongs to. It usually manages. It fails around family history, when a relative's condition is discussed, and around interpreters, parents and carers, where the person speaking is not the person being described.

The example error, "history of type 2 diabetes", would put a condition in a patient's record that follows him to insurance medicals and future clinicians. Other common versions: a parent's account of a child's symptoms recorded as the child's own words; a daughter translating, with her opinion recorded as the patient's; a doctor's explanation ("this can cause heart problems") recorded as the patient's history.

How to catch it

Scan the past medical history, family history, medicines and social history for items that were mentioned but belong to someone else. Anything in these sections that you do not recall the patient saying about themselves needs a second look. Ask whether the history lines were things you asked or things you explained.

How to reduce it

Name the person when you move topics: "Your wife has diabetes; do you?" Introduce a third party at the start: "This is the patient's daughter, who is helping with the history." With any interpreter, check your scribe's rules for multilingual sessions first, since accuracy varies widely and some tools handle a second language much worse than the first.

6. Template carry-over: yesterday's sentences in today's note

This one is not the model's invention but yours. Templates, macros and "copy previous note" features pre-fill text so the scribe only has to fill gaps. If a field says "Non-smoker. No known drug allergies." and nothing was discussed, the sentence stays, and you have attested to a conversation that did not take place.

A subtler version is a custom prompt that tells the scribe to "include smoking, alcohol and allergy status". When the topics did not come up, a good model should leave the field blank; a weaker setup will fill it with a default or with the previous record's content if it has access to one.

How to catch it

Look for any sentence that is true-sounding, generic and about something the patient has not recently discussed. Social history, allergies, medication lists and "no other concerns" are the usual homes. Ask "was this said today?" If not, it is a carry-over, even if it is probably still true.

How to reduce it

Set templates so unspoken fields read "not discussed" instead of a default value. Keep copy-forward fields, such as long-term conditions or allergies, visually separate and labelled with the date last confirmed. A good habit is to confirm allergies aloud once per visit; it takes five seconds and gives the scribe a sentence to write.

A two-minute reading order that catches all six

Instead of reading top to bottom, read in the order of risk. This is a routine you can learn in a week, and it works on any tool:

1. Plan and medicines (30 seconds). Every drug, dose, test, referral and follow-up interval. Compare with what you ordered. Catches wrong dose and omissions.
2. Sides, numbers and units (20 seconds). Left and right, dates, measurements, durations. Catches wrong side and invented readings.
3. Negatives and red flags (20 seconds). Each serious symptom, present or absent. Catches negation flips.
4. Examination (20 seconds). Each finding: did I do it and say it? Catches invented findings.
5. History sections (20 seconds). Whose condition is this, and was it discussed today? Catches speaker mix-ups and carry-over.
6. Last question (10 seconds). Is there anything I said at the door or at the end that isn't here? Catches omissions.

That is about two minutes. On a day of 20 patients it is 40 minutes of checking, which is why a scribe only saves time if the drafts are good and the template is tight. See whether an AI scribe is worth it for a small private clinic for the arithmetic of time saved against time checked.

Scoring your own notes for a fortnight

You will not know which errors your tool favours until you count them. For two weeks, keep a tally on a sheet of paper or a spreadsheet. After signing each note, tick the type of any error you changed. Illustratively, a solo practitioner might see this after 80 consultations:

These are made-up numbers; your own tally will have its own shape. Read it in two directions. Frequent, usually low-harm errors (here, omissions and carry-over) point at the template and your closing summary. Rare, high-harm ones (here, wrong dose and negation flips) decide where you slow down. Rare errors still deserve a dedicated check because the cost of one is out of proportion to its frequency.

Prompt and template wording that gives the scribe less rope

If your tool allows custom instructions, a few lines change the failure profile. Here is a wording you can adapt, written for a general consultation template:

Write the note only from what was said in the consultation.
If a section was not discussed, write "Not discussed" and nothing else.
Do not write examination findings, observations or diagnoses unless the clinician said them aloud.
Record negatives ("denies", "no") exactly as stated and list them separately under "Pertinent negatives".
Attribute every history item to the person who said it. If a family member's condition is mentioned, label it as family history.
Copy doses, strengths, sides and numbers exactly as spoken. If a number or side is unclear, write [CHECK] in its place.
Do not carry over any text from earlier notes.

The [CHECK] marker is the most useful line. A draft that says "ramipril [CHECK] mg" forces you to look; a draft that confidently says 15 does not. Test the instructions on five old, uncontroversial consultations before trusting them, and note whether the tool actually obeys them. Some tools ignore custom instructions in places.

A sample of the output you might get back for the knee example, labelled as illustrative:

Presenting complaint: Left knee pain, three weeks.
Pertinent negatives: Chest pain, breathlessness.
Medicines: Ramipril 5 mg daily. Plan to increase to 10 mg [CHECK].
Examination: Not discussed.
Family history: Wife has diabetes.
Plan: Left knee X-ray. Review in two weeks.

What you would fix: confirm 10 mg and keep it, write the examination you did, add the kidney bloods, and add the safety-netting advice. The draft is useful, but it is a draft.

When a tool makes checking easier or harder

Judge scribes less by their headline accuracy and more by how quickly you can audit them. Four features matter:

• Linking a sentence to the transcript. Click a line, see the words that produced it. This turns an invented finding from a memory test into a ten-second lookup.
• Flags for low confidence. A marker on numbers and doses that the model heard poorly.
• Section control. The ability to leave a field blank rather than force one.
• Edit tracking. A record of what you changed, useful for your own error tally.

Heidi, for instance, lists a free plan with unlimited ambient documentation and a trial-limited allowance for other features, and a paid clinician plan at $110 per user per month billed annually; Jane App lists its scribe add-on at $15 a month per opted-in practitioner. Prices change, so check the vendor page. The point is that a cheap tool with poor traceability can cost more in checking time than an expensive one with better auditing. For a fuller price comparison, see AI scribe costs compared for small clinics.

Consent, storage and what a transcript means for your records

Error checking links directly to a data question: if you keep the transcript or audio so you can verify a doubtful line, you now hold a second record containing sensitive information. Decide the policy before you start. Does the vendor keep audio, for how long, and where? Is there a business associate agreement or data processing agreement that covers it? Do your patients know? The tutorial on rolling out an AI scribe without losing patient trust covers wording for consent.

The record-keeping rules for your profession decide whether a transcript counts as part of the clinical record. Ask your indemnifier or professional body rather than guessing. And do not paste a consultation into a free, consumer chatbot to "tidy" it; that is a different data flow from a scribe with a signed agreement.

Seven short scenarios to test your own tool

Before you trust a scribe with real patients, run it through deliberately awkward practice conversations, perhaps with a colleague playing the patient. Each of these probes one error type:

1. The door handle. Add a new complaint as you stand to leave. Does it reach the note?
2. The silent examination. Examine without speaking results. Does the note invent any?
3. Similar numbers. Mention 15 mg and 50 mg in the same minute and prescribe one. Which is written?
4. Both knees. Discuss the right knee as the comparison for a left-knee problem. Is the side right?
5. The denied symptom. Ask about chest pain and have the patient deny it quietly. Does the "no" survive?
6. The relative's illness. Have the patient describe a parent's heart condition. Where does it land?
7. The empty social history. Say nothing about smoking. Does a smoking status appear?

Keep the results. If the tool fails a test, adjust the template or your speech habits and run it again. Repeat the set after any vendor update, since a model change can shift behaviour without warning. This is an afternoon of work that tells you more than any vendor's accuracy claim.

What to do when you find an error after signing

It will happen. Handle it as you would any record error: do not overwrite or delete the original; add a dated, signed addendum that states what was wrong and what is correct, and, if the error could have affected care (a wrong dose communicated, a missed referral), contact the patient and act on it. Check with your indemnifier about how they want incidents logged. Then ask why it got through: which of the six types was it, and which step of your reading order should have caught it? Adjust the routine, not just the note.

Once notes are under control, the next admin job is usually letters and the inbox, covered in using an AI secretary for referral letters, inbox sorting and insurer forms. If you would like to map your reading order and template with someone, that is what my AI implementation consultation covers, though everything above works without it.

Questions doctors ask about checking scribe notes

How long should checking an AI scribe note take?

Aim for two to three minutes on a routine consultation once you have a fixed reading order. Read the plan and medicines first, then negatives and sides, then the history. If checking regularly takes longer than writing the note yourself would have, the template or the tool needs changing, not your reading speed.

Does a better scribe remove these errors?

No. Vendors improve accuracy, but every tool that turns speech into prose can omit, invent or misattribute. Treat error types as permanent features of the technology and keep a review step, whatever the product. Compare tools on how easy they make checking, such as linking sentences to the transcript.

Should I keep the audio or transcript to check against?

Only if your tool, your consent wording and your records policy allow it, and you know where it is stored and for how long. Some vendors delete the audio as soon as the transcript exists and keep the transcript only for a short review window. Decide this before rollout, and tell patients what happens to the recording.

Who is responsible for an error in a signed note?

The clinician who signs it. A scribe is a drafting aid, and your signature says you have read and agree with the content. Ask your indemnifier or professional body how they view ambient scribes, since guidance differs by profession and jurisdiction.

Can I let a colleague or assistant sign off the draft?

Not as a general rule. The person who held the consultation is the only one who knows what was said and found. An assistant can fix formatting or fill in administrative fields, but the clinical content needs the clinician who saw the patient.

Further reads

• What Is an AI Scribe and How Does It Work in a Consultation? — How a scribe turns a consultation into a note, step by step.
• AI Answering Service for Doctors: What It Costs and What You Get — What an AI answering service costs and what you get.
• AI Dictation Software for Doctors: Dictation or Ambient Notes? — Dictation versus ambient capture, and which errors each produces.
• AI-Generated SOAP Notes for Chiropractors: A Review Checklist — A SOAP-note review checklist from a neighbouring specialty.
• AI Chatbot for a Doctor's Website: Scope, Disclosure, Handover — Another patient-facing AI job, with scope and handover rules.
• AI Claim Scrubber: What It Catches and What It Misses — Modifier, bundling and diagnosis edits are caught; eligibility, documentation and fresh payer rules often are not. Test one on 50 claims.
• AI Medical Coding Software: Audit Trails and Upcoding Risk — Suggest or submit? Who signs, what the audit trail must record, how to spot upcoding drift and what to test on 100 old charts.
• AI Medical Transcription Services vs Human Transcribers for Small Clinics — Per-minute quotes, turnaround, BAA and an error-check step, plus a five-dictation test that shows which service suits your clinic.
• AI Platform for Doctors: One Suite or Separate Scribe and Phone Tools? — A buying rule for small practices: when one suite wins, when separate scribe and phone tools win, and the integration test to run first.
• AI Scribe Security: Audio, Retention and BAA Vendor Questions — A contract-ready list of scribe security questions, evidence requests and a weighted vendor scoring sheet.
• AI Scribe for Dentists: Will It Write Back to Your Software? — A dental scribe comparison that starts with the saved clinical note, not a vendor integration badge.
• AI Scribe for GPs: Costs, Free Tiers and the NHS Pilot Limit — A GP scribe budget with free and paid options, hidden review time and the four-month pilot boundary.
• AI Scribe for Medical Practices: Pick by Country and EHR — A practical decision matrix for choosing a clinical scribe by jurisdiction, EHR connection and specialty fit, with a scored pilot.
• Dermatology EHR With AI: What Is Real for a Small Clinic — Dermatology EHRs sorted into real AI (ambient notes, code suggestions) and marketing (diagnosis from photos), with a cost sum for a small clinic.
• AI Scribe Consent Form: US Wording and State Recording Laws — Template wording for an AI scribe consent form, a plain map of one-party and all-party recording states, and an opt-out that still gets the patient seen.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Heidi pricing page and Jane pricing page (both fetched 4 November 2026); SimplePractice Note Taker transcript-retention FAQ (fetched 4 November 2026).