It is 6:40 in the evening. A hospitalist has eleven notes left to sign and a partner texting that dinner is cold. The ambient scribe has drafted every one of them, and each is a clean, complete, professional SOAP note. She could click the signature button eleven times in under a minute, and no one would stop her. The moment her finger hovers over that button is the most important moment in the entire encounter, because the instant she signs, a machine-drafted paragraph stops being a suggestion and becomes the legal record, her legal record, indistinguishable from anything she ever wrote herself. This lesson is about what has to happen in the seconds before that click.

The Signature Changes Everything

Start with the single fact that reorganizes how you should think about an AI-drafted note: the signature is a transformation, not a formality. Before you sign, the draft is a proposal from a machine, an artifact with no authority and no accountability attached. After you sign, it is your professional attestation that this is an accurate account of the care you provided. The record does not say "drafted by AI, lightly reviewed." It says, in effect, "I, the licensed clinician, affirm this." The law, the coder, the next clinician, the auditor, and, if it ever comes to it, the plaintiff's attorney all read a signed note as your words and your judgment. The ambient scribe that produced the first draft is nowhere in that chain of accountability. It cannot be sued, sanctioned, or called to testify. You can. That asymmetry is the whole reason verification is not optional.

This is worth stating plainly because the convenience of a good scribe quietly encourages the opposite belief. When a tool is fast, polished, and usually right, it feels like a colleague who has done the work for you, and signing feels like a courtesy acknowledgment rather than an act of authorship. It is not. Signing is authorship. The most useful reframe you can adopt is this: the scribe wrote a draft, but you are the author, and an author who signs work they did not read is accountable for every word in it anyway. The discipline this lesson teaches, reading before you sign, is simply the behavior that makes the authorship real rather than fictional.

Understand the machine you are signing behind, because knowing where it fails tells you where to look. An ambient scribe runs in two stages, and each stage fails differently. The first stage is transcription: a speech-to-text model turns the audio of the visit into words. This is where soundalike errors are born. "Hypertension" and "hypotension" differ by a single vowel and mean opposite things; "fifteen" and "fifty" are one clipped syllable apart; a mumbled "losartan" becomes "lorazepam," an antihypertensive swapped for a benzodiazepine. The second stage is generation: a language model takes that raw transcript, plus context, and composes it into the tidy SOAP note you actually read. This is where fabrication is born, because the generative stage is rewarded for producing something that reads like a complete, well-formed note, and a complete note has pertinent negatives, a coherent assessment, and a clean plan, whether or not the visit actually supplied them. When the transcript is thin, the model does not leave a blank. It fills the blank with the most plausible clinical prose, and plausible is not the same as true. Knowing this, you can predict the errors before you see them: garbled numbers and drug names from stage one, invented completeness from stage two. Your eyes should go to exactly those places.

The Read-Before-You-Sign Discipline

The core skill is easy to name and hard to hold under pressure: you read the note before you sign it, every time, with attention proportioned to what could go wrong. Notice the phrase "proportioned to what could go wrong." Reading before you sign does not mean re-deriving the entire encounter from memory or treating every note as a research project. It means directing your limited attention, on a busy shift, to the parts of the note where an uncaught error would actually hurt a patient or expose you. A misplaced comma in the social history is not where you spend your scarce seconds. A wrong dose in the plan is exactly where you spend them. The skill is triage: knowing which lines carry the risk and reading those lines like your name is on them, because it is.

The reason this must be a fixed discipline rather than a good intention is the lesson you already learned about automation bias. A good scribe trains you, one accurate note at a time, to stop checking. The mechanism is quietly corrosive: each accurate note is a tiny reward for trusting, and over a few hundred notes your brain learns, without ever deciding to, that reading is wasted effort. This is not a character flaw. It is exactly how expertise is supposed to work. A resident who has drawn ten thousand uneventful blood cultures stops treating each one as a novel event, and that automaticity is what lets an experienced clinician move fast. The problem is that the scribe earns your trust on the easy ninety-five notes and then spends it on the hard five, and the five where it fails are not announced. They look identical to the ninety-five where it did not. The only defense that survives a hard shift is a rule that does not depend on how you feel in the moment: I read the high-stakes elements of every AI note before I sign, regardless of how reliable the tool has been and how late it is. A rule fires whether or not your vigilance is intact. An intention evaporates at 6:40 with dinner cold. Build the rule.

There is a second reason to make it a rule and not a judgment call, and it is about how the errors distribute. If the scribe made large, obvious errors, you would not need a discipline; the errors would catch your eye on their own, the way a note that suddenly describes the wrong patient's chief complaint would. The dangerous errors are precisely the ones calibrated to slip through: a dose that is plausible for the drug but wrong for this patient, a laterality that is internally consistent throughout the note because the model committed to the wrong side early and stayed loyal to it, a negative that fits the clinical picture so well you would have written it yourself. These errors are invisible to a glance and visible only to a check. A glance asks "does this look like a reasonable note," and the answer is almost always yes, because producing reasonable-looking notes is the one thing the generation stage is guaranteed to do well. A check asks "is this specific high-stakes element actually what happened," which is a different question entirely, and the only one that catches the calibrated error.

The scribe wrote a draft. You are the author. An author who signs work they did not read is accountable for every word in it anyway. Reading before you sign is what makes your authorship real.

What to Check First: The High-Stakes Elements

Because attention is finite, you check in order of harm. The elements below are the ones where an AI error most reliably reaches a patient or an auditor, roughly in the order you should verify them. Learn this order until it is automatic, so that even on the worst shift your eyes go to the dangerous places first. The order is not arbitrary. It tracks two things at once: how fast an error can hurt a patient, and how likely the pipeline is to produce that particular error. Medications and doses top the list because they are both the fastest to harm and, being numbers and soundalike names, among the most likely to be garbled by transcription. Laterality follows because a flipped side is quietly catastrophic and easy for either stage to invert. Pertinent negatives come next because they are the signature failure of the generative stage. And the assessment and plan close the list because, while less likely to contain a raw factual error, they are what actually steers the next clinician's hand. Internalize the reasoning, not just the sequence, and you can adapt the order to your own specialty where the harms differ.

Medications, Allergies, and Doses

Nothing else on the note can hurt a patient as fast as a wrong medication, a missing allergy, or a wrong dose, and these are exactly the elements the transcription stage most reliably garbles, because they are numbers and soundalike names. Check every medication in the plan against what you actually prescribed. Check the dose, the route, and the frequency, not just the drug. A right drug at a wrong route can be as dangerous as a wrong drug: vincristine given intrathecally instead of intravenously is a classic, fatal example, and while your ambient scribe is not ordering chemotherapy, the principle holds down to the ordinary insulin whose "units subcutaneous" quietly became "units" with no route, or the antibiotic whose "every eight hours" was transcribed as "every eight days." Confirm that documented allergies match the truth and that no allergy has silently dropped out; a dropped penicillin allergy does not merely lose information, it removes the guardrail that every downstream prescriber and every order-entry alert depends on, so the next clinician who orders a cephalosporin does it blind. A note that reads beautifully and lists the wrong dose of an anticoagulant is not a small error; it is the kind of error that ends up in a mortality review, and a warfarin or a direct oral anticoagulant at the wrong dose can bleed a patient or clot them within days. This is the first place your eyes go, every time, without exception.

Laterality and Anatomic Specifics

Left versus right, upper versus lower, proximal versus distal: laterality is a small word that carries enormous consequence, and it is easy for both the transcription and generation stages to flip. A note that documents the wrong side can drive a wrong-site procedure, a wrong imaging order, or a misdirected referral. Wherever the note names a side or a specific anatomic location, stop and confirm it against reality. This is a two-second check with a catastrophic downside if skipped.

Pertinent Negatives

A pertinent negative is a symptom the patient does not have that matters precisely because it shapes the differential: "denies chest pain," "no fever," "no focal weakness." These are among the most dangerous elements in an AI note for two opposite reasons. The generation stage can add negatives that were never actually elicited, padding the note with an authoritative "denies" for a question no one asked. And it can drop or flip a negative that was reported, turning a patient who reported chest pain into one who denied it, or vice versa. Both failures corrupt the clinical reasoning that later readers will trust. Verify that every documented negative reflects what actually happened in the room, and that a negative you relied on to rule something out is truly present and truly correct.

The Assessment and Plan

Finally, read the assessment and plan as if a covering colleague will act on them tonight, because one might. This is the part of the note that drives what happens next: the medications continued, the follow-up ordered, the diagnosis carried forward. The generation stage can overstate certainty here, rendering a tentative "possibly early heart failure, will monitor" as a confident diagnostic label the record then treats as settled. It can also carry forward a plan element from a template that does not match tonight's decision. Read the plan and ask whether it is what you actually decided, at the level of certainty you actually hold. The assessment and plan are where the note stops describing the visit and starts steering the future.

A Worked Example: Ten Seconds That Change the Record

Watch the discipline turn a dangerous draft into a defensible note. A nurse practitioner sees a patient for a diabetic foot ulcer on the left heel. The ambient scribe produces a clean note. Under the surface, three things went wrong: the plan reads "continue metformin 1000 mg twice daily" when she reduced it to 500 mg twice daily because of worsening renal function; the exam documents the ulcer on the "right heel"; and the review of systems includes "denies fever and chills," which she never asked, and which matters, because an undocumented fever would change her concern for underlying osteomyelitis.

Now watch an author, not a rubber stamp, read it. Her eyes go first to the medications: she catches the metformin dose immediately, because she checks every dose against what she decided, and she corrects 1000 to 500. Her eyes go to laterality: she catches "right heel," a wrong-site error that could have sent imaging and a podiatry referral to the wrong foot, and corrects it to left. She reads the pertinent negatives and stops at "denies fever and chills," because she does not remember asking; she deletes the fabricated negative and, realizing she genuinely does not know the patient's fever history, steps back in to ask, discovering the patient did have chills two nights ago, a finding that changes her plan. Total added time: under a minute. What that minute prevented: a renal-dosing error, a wrong-site workup, and a missed sign of a limb-threatening infection. That is the entire value proposition of reading before you sign, compressed into one note.

Notice what made the difference. It was not that she was smarter or more caring than a clinician who would have signed the draft. It was that she had a fixed order of checks that fired automatically: meds, laterality, negatives, plan. The draft was identical for both clinicians. The signature was not. That is the humbling and empowering truth at the center of this lesson: the tool does not decide whether the record is safe. The person hovering over the button does, and they decide it with a habit, not with heroics.

It is worth sitting with how ordinary each of those three errors was, because none of them looked like a malfunction. The metformin line was not garbled or nonsensical; it was a perfectly formatted, clinically reasonable dose that happened to be the dose from a previous visit rather than tonight's decision, the kind of stale carry-forward that a template-aware model produces effortlessly. The "right heel" was internally consistent, appearing in the exam and echoed in the plan, because once the generation stage committed to a side it stayed loyal to it, which is exactly what makes a laterality flip so treacherous: the note does not contradict itself, so nothing inside the note flags the error. The fabricated "denies fever and chills" was the most seductive of the three, because it was clinically appropriate. A careful clinician would have asked about fever in a diabetic foot ulcer, so the model, trained on thousands of such notes, supplied the negative that a complete note "should" contain. It was not a random hallucination. It was a plausible inference presented as an observed fact, and that is the signature failure mode you are guarding against. The draft was not broken. It was fluent, and fluent is precisely what hides the error.

How to Actually Edit, Not Just Read

Reading is only half the discipline. When you catch an error, you have to correct it well, and there is a right and a wrong way to do that. The wrong way is to leave the false content in place and add a contradicting line elsewhere, so the note now says two incompatible things and a later reader cannot tell which is true. The right way is to make the note say what actually happened: delete the fabricated negative rather than annotating around it, change the wrong dose to the right one, fix the side, rewrite the overconfident assessment to match your real certainty. A clinical note is not a debate transcript that preserves every claim; it is a factual record, and your job when editing is to make every line true, not to accumulate corrections beside the errors.

Two habits make this reliable. First, when you delete a fabricated finding, pause to ask whether the underlying information actually exists. The nurse practitioner in the earlier example did not just delete "denies fever and chills"; she recognized that removing it left a real gap in her knowledge and went back to fill it. Deleting a fabricated fact can reveal that you never actually have the real one, and that is a clinical prompt, not just an editing task. Second, when you correct something material, make sure the correction propagates. If you change a dose in the plan, confirm the corresponding order is right; if you fix laterality, confirm the imaging request and the referral point to the correct side. A note lives inside a web of orders and communications, and a correction that stops at the note while the wrong order sails onward has fixed the record but not the care.

A Mental Model: Draft, Then Author

The cleanest way to hold all of this is a two-word sequence: the machine drafts, then you author. Drafting is assembling raw material into a shape. Authoring is taking responsibility for every claim in the finished piece. A journalist may work from a researcher's notes, but the byline, and the accountability for every fact under it, belongs to the writer who signed it. The ambient scribe is your researcher. It is fast, tireless, and genuinely useful, and it does not know what is true. You are the writer whose name goes on the piece. The read-before-you-sign discipline is simply the moment where drafting becomes authoring, and skipping it does not remove your byline. It only means you put your name on words you never actually read.

Attestation: You Own What You Sign

The formal name for what your signature does is attestation. When you attest to a note, you are formally affirming that its contents are accurate and reflect the care you provided and the clinical judgment you exercised. Attestation is not weakened by the fact that a machine wrote the first draft, any more than it would be weakened if a human scribe or a resident had drafted it. In every one of those cases, the accountability for the signed note is yours. This is not a harsh new rule invented for AI; it is the oldest rule in medical documentation, applied to a new drafting tool. A note has always meant "the signing clinician stands behind this." AI does not change the meaning of your signature. It only changes how easy it has become to sign something you did not actually author, which is precisely why the read-before-you-sign discipline matters more now, not less.

Attestation also has a documentation dimension worth naming, because the standard of care around AI is evolving in a direction that rewards a specific habit. When you disagree with something the draft asserted and change it, a single clause explaining why you made the change materially strengthens the record. "Scribe draft listed metformin 1000 mg; corrected to 500 mg for rising creatinine" is not defensive clutter. It is the difference between a chart that shows a clinician thinking and a chart that shows a clinician clicking. If the case is ever reviewed by a coder, a RAC auditor, a Joint Commission surveyor, or a plaintiff's attorney, the note that documents the clinician's reasoning is the note that reads as care. You do not need to annotate every trivial edit. You do need to leave a trace when your judgment overrode the draft on something that mattered, because that trace is the visible proof that a human, not a machine, made the decision that reached the patient. Some state disclosure laws are pushing in a parallel direction: when generative AI is used in patient communications or in diagnosis and treatment, patients may be entitled to know, and a note that a licensed clinician read and reviewed the AI output is often exactly what moves a communication out of the disclosure requirement and into the ordinary standard of documented care.

There is a constructive flip side worth holding onto, so this does not read as pure warning. The same signature that carries your accountability also carries your protection, when you use it well. A note you actually read and corrected, where the dose is right, the side is right, the negatives are true, and the plan is what you decided, is a strong, defensible record that will hold up to any later reader. The discipline does not just prevent harm; it produces a better record than most clinicians ever had time to write by hand, because the scribe did the drafting labor and you did the verifying judgment. Used this way, the ambient scribe plus a disciplined author is genuinely better than either alone. That is the goal: not fear of the tool, but mastery of the handoff between what it drafts and what you sign. Fear makes clinicians reject a genuinely helpful tool; blind trust makes them sign whatever it produces. Mastery is the narrow, correct path between the two, and it is built from exactly one repeatable behavior performed on every note, no matter how tired you are or how good the tool has been.

Key Takeaways

• The signature is a transformation, not a formality: before you sign, an AI draft is a machine's proposal with no accountability; after you sign, it is your legal record, read by every later clinician, coder, auditor, and attorney as your words and your judgment.
• The scribe cannot be sued, sanctioned, or called to testify; you can. That asymmetry is the whole reason verification is not optional. The scribe wrote a draft, but you are the author.
• Read before you sign, every time, with attention proportioned to what could go wrong; the skill is triage, directing your scarce seconds to the lines that carry the risk, not re-deriving the whole encounter.
• Make it a fixed rule, not a good intention, because a good scribe trains you to stop checking; a rule fires whether or not your vigilance survives a hard shift, an intention evaporates at 6:40 with dinner cold.
• Check in order of harm: medications, allergies, and doses first (numbers and soundalike names the transcription stage garbles), then laterality and anatomic specifics, then pertinent negatives (added, dropped, or flipped by the generation stage), then the assessment and plan.
• Pertinent negatives are doubly dangerous: the model can add a "denies" no one asked, and can drop or flip a negative that was reported; both corrupt the clinical reasoning later readers trust.
• Read the assessment and plan as if a covering colleague will act on them tonight; the generation stage can overstate certainty or carry forward a template element that does not match what you actually decided.
• Attestation means you own what you sign, exactly as you always have; AI does not change the meaning of your signature, it only makes it easier to sign something you did not author. A note you actually read and corrected is a stronger, more defensible record than most clinicians ever had time to write by hand.