Two summaries land on a nurse's screen an hour apart. The first invents a fact: it lists an aspirin allergy the patient has never had, pulled from nowhere the chart contains. The second hides one: it silently omits the patient's documented penicillin allergy, which is right there in the record but never made it into the summary. Both summaries are wrong. But they are wrong in opposite directions, and the tragedy is that the habits that catch the first are useless against the second. The invented allergy she might notice by reading closely. The missing allergy she will never catch by reading the summary at all, no matter how carefully, because it is not on the page to be read. This lesson is about the two failure modes summaries share, why they demand different defenses, and the one fast cross-check that catches both. It is the practical core of everything the chapter has taught: once you can name how a summary goes wrong and check for it in seconds, the summary becomes a tool you can trust exactly as far as it deserves, and no further.

The Two Failure Modes: Adding and Hiding

Every summary can fail in exactly two ways, and naming them precisely is the whole foundation of catching them. The first is fabrication: the summary adds what is not there. A diagnosis the patient never carried, a medication they never took, a lab value never drawn, an allergy that does not exist. The information appears on the page as a confident assertion, but the source contains no basis for it. The model, generating fluent text, produced a plausible fact that this particular chart does not support. Fabrication is the failure of addition.

The second is omission: the summary hides what is there. A real allergy left out. An active diagnosis dropped. An abnormal value compressed into silence. A critical medication that simply did not survive the summarization. The information exists in the source, correct and important, but never appears in the summary. The model, compressing toward a shorter narrative, discarded something that mattered. Omission is the failure of subtraction. Adding what is not there and hiding what is: these are the two, and only two, ways a summary betrays you.

It is worth pausing on why these are the complete set, because clinicians sometimes imagine a third category, the summary that gets a fact subtly wrong, a value transcribed as 4.1 instead of 1.4, a dose stated as twice daily instead of once. But look closely and this is not a third mode; it is both of the first two at once. The wrong value is a fabrication (the 4.1 that does not exist in the source) layered on an omission (the 1.4 that does but was dropped). Every summary error decomposes into adding, hiding, or both. That is not a pedantic point. It means that if your verification reliably catches both addition and subtraction, it catches everything a summary can do to you, which is exactly why the bidirectional cross-check is not one defense among many but the complete defense.

A summary fails in two directions: it adds what is not there, or it hides what is. One you can catch by reading. The other you can only catch by leaving the summary and returning to the source.

Why They Need Opposite Defenses

The reason these two matter so much as a pair is that they are asymmetric in the most consequential way possible: they are caught by opposite methods, and the method that catches one is blind to the other. Fabrication is, in principle, catchable by reading the summary, because the false fact is present on the page. A confident but surprising claim can trip your clinical pattern recognition; a stated allergy or diagnosis can be checked against the source to confirm it is real. Reading the summary critically is a genuine, if imperfect, defense against fabrication, because there is something on the page to be skeptical of.

Omission is the opposite, and this is the single most important idea in the lesson: omission cannot be caught by reading the summary harder. A dropped penicillin allergy leaves no trace on the page. There is no sentence to be skeptical of, no odd claim to snag on, no shape where the fact used to be. You can read the summary ten times with perfect attention and never detect the absence, because attention operates on what is present, and an omission is defined by what is not. The only way to catch an omission is to leave the summary entirely and go check the source for what should be there. No amount of scrutiny applied to the summary itself will ever reveal what the summary silently left out.

This asymmetry has a direct practical consequence that clinicians routinely get wrong. The instinct, when told a summary might be unreliable, is to read it more carefully. That instinct is half right and half dangerous. It helps against fabrication and does nothing against omission, and omission is both the more common failure and the one more likely to hide the fact that hurts the patient. Careful reading of the summary feels like diligence, and against half the failure modes it is diligence, but it produces a false confidence that you have checked, when the more dangerous failure sailed straight through your careful reading untouched. The feeling of having read carefully is not the same as having verified.

This confusion between reading and verifying deserves to be named directly, because it is where competent, conscientious clinicians get caught. Verification is not an intensity of attention; it is a direction of attention. Reading the summary, however carefully, points your attention at the summary, and the summary is precisely the artifact that cannot tell you what it omitted. Verification points your attention at the source, which is the only place the omission is visible. A clinician who reads the summary three times is applying more of the wrong kind of attention. A clinician who glances once at the source allergy list is applying the right kind. Effort is not the variable that matters. Direction is. And because omission is the more dangerous and more common failure, the direction that matters most is the one that points away from the summary and back to the record.

The Fast Cross-Check Habit

Because reading the summary cannot catch omission, the defense has to be a positive check against the source: a short, fixed list of the things you actively confirm are present and correct, rather than passively waiting for a problem to announce itself. The habit is simple and fast, and it targets exactly the fact classes where a fabricated or dropped item does clinical harm. Before you rely on a summary for a decision, spot-check four things against the source: the abnormal values, the active problems, the medications, and the allergies.

Notice what this check does that reading cannot. For fabrication, you are confirming that each stated item is actually supported by the source: is this listed allergy real, is this diagnosis really in the record, was this value actually drawn. For omission, you are doing the harder and more important thing, confirming that the source's items all made it into your understanding: are all the documented allergies accounted for, are the active problems complete, is any current medication missing. That second direction, checking the source for what the summary might have dropped, is the only thing that catches omission, and it is precisely the check that reading the summary alone can never perform. The cross-check is bidirectional by design: summary-to-source catches fabrication, source-to-summary catches omission.

There is a discipline that keeps the check honest, and it is worth stating plainly: you name the four classes out loud, in the same order, every time, so the habit does not quietly shrink to the two you happen to remember on a tired night. Clinicians who run the check informally tend to drift toward allergies and medications, the two that feel most urgent, and let abnormal values and active problems slip. That drift is exactly how a dropped potassium of 6.8 or a quietly deleted diagnosis of heart failure survives the check. A fixed, spoken list, abnormal values, active problems, medications, allergies, is not bureaucratic ritual. It is the mechanism that keeps the more forgettable and more silent failures inside the net. The list you say every time is the list that protects you on the night you are too busy to think.

Why These Four Fact Classes

These four are not arbitrary. They are the fact classes where a wrong or missing item translates most directly into patient harm. A fabricated or dropped allergy drives a dangerous order. A wrong or missing medication is a prescribing error waiting to happen. A dropped abnormal value hides the thing that should change management. A dropped active problem is a diagnosis quietly falling out of the care plan. Other facts can be wrong with less consequence, and you cannot re-verify everything in a summary without defeating its purpose, since the whole point of the summary is to save time and a check that re-reads the entire chart would give that time straight back. So you concentrate your fast check on the four classes that carry the harm, and you do it every time you are about to act on the summary, not just when something feels off, because a dangerous omission never feels off and never announces itself.

There is a reason these four map so cleanly onto the harm, and it connects this lesson to the medication-reconciliation and allergy-verification disciplines you already practice without a computer. Long before AI summaries existed, the safety culture of medicine had already identified allergies, medications, active problems, and critical results as the facts you confirm directly rather than take on faith, precisely because a wrong or missing one of these is how patients get hurt. The AI summary does not change that list; it just adds a new way for those facts to be wrong or missing. So the four-fact cross-check is not a novel burden invented for AI. It is the extension of an existing clinical instinct, medication reconciliation, allergy confirmation, results review, to a new source of error. That is why it should feel familiar rather than foreign, and why it is defensible to a surveyor: you are applying an established standard of care to a new tool, not inventing a special exception for AI.

A Worked Example: The Two Wrong Summaries

Return to the nurse and her two summaries. The first, with the fabricated aspirin allergy, is the easier catch. As she does her cross-check, she confirms allergies against the source and finds no aspirin allergy documented anywhere. The summary added a fact the record does not support. She removes it, and no harm follows, because she checked the stated allergy against the source rather than trusting it. A clinician reading critically might even have paused at the aspirin allergy on the page, since the fabrication is present to be doubted.

The second summary is the dangerous one, and it shows exactly why reading is not enough. It looks clean. It has no odd claims, nothing to trip a careful reader, no aspirin allergy or surprising diagnosis to pause at. It reads as a perfectly reasonable, unremarkable summary, which is precisely the problem: it silently dropped the patient's documented penicillin allergy. A nurse who reads this summary carefully, even brilliantly, finds nothing wrong, because nothing wrong is on the page. If she relies on it and the patient is ordered a penicillin-class antibiotic, the omission becomes a potentially fatal allergic reaction. The only thing that catches it is the cross-check: she goes to the source allergy list, sees the documented penicillin allergy, notices it is absent from the summary, and stops the error before it happens. Reading gave her nothing. The source gave her everything.

Set the two side by side and the lesson is unmistakable. The fabricated allergy was caught because a false fact was present to be checked. The dropped allergy was caught only because she checked the source for what should be present, independent of the summary. Same nurse, same four-fact habit, two opposite failure modes, both caught, because the habit checks in both directions. Had she relied on reading alone, she would have caught the harmless fabrication and missed the dangerous omission, which is the exact inversion of what safety requires.

Dwell on that inversion, because it is the heart of why this matters. If a verification method reliably caught the harmless error and reliably missed the catastrophic one, you would call that method worse than useless: it would consume effort, generate a feeling of safety, and leave the real danger untouched. Yet that is exactly the profile of "just read the summary carefully" as a defense. It catches the fabricated aspirin allergy, which was never going to kill anyone, and it misses the dropped penicillin allergy, which could. A defense that is strong against the trivial failure and blind to the lethal one is not a partial defense; in the way that counts, it is an anti-defense, because the false confidence it produces actively discourages the source check that would have caught the real thing. This is why the lesson insists so hard that reading is not verification. The gap is not a matter of degree. It is the difference between checking the thing that can hurt the patient and checking the thing that cannot.

Verifying a Summary Against the Source Record

The two-allergy story is clean because allergies are a short, discrete list. Most of the summaries you act on are messier: a hospitalist's overnight sign-out, a discharge summary compressing a five-day admission, an AI-generated recap of a specialist consult. To see how verification works on a real summary, watch one abnormal value get dropped, then watch it get caught. Consider a patient admitted for community-acquired pneumonia, improving on day three. The source record from the overnight labs contains, among the usual chemistry, a potassium of 6.1, freshly resulted and not yet addressed, alongside a normal sodium, a stable creatinine, and a white count trending down. The AI summary handed to the day team reads beautifully:

Day 3, community-acquired pneumonia. Afebrile 18 hours, oxygen weaned to room air, white count improving. Labs reviewed and largely reassuring, chemistry stable. Plan: continue oral antibiotics, anticipate discharge tomorrow.

Read that summary as a clinician and nothing snags. It is fluent, it is plausible, it matches the trajectory of a patient getting better, and the phrase "chemistry stable" is the kind of thing an experienced writer produces at the end of a good admission. That is precisely the trap. "Chemistry stable" is not a lab value; it is a summary of lab values, and it silently swallowed the one result that should have stopped the discharge conversation cold. A potassium of 6.1 is not stable chemistry. It is a critical value that demands an ECG, a repeat draw, and treatment before anyone talks about going home. The summary did not lie in the sense of stating a false number. It committed the more dangerous sin: it compressed a critical abnormal value into a reassuring generality, and the abnormality vanished into the word "stable."

Now watch the cross-check work. The day clinician, before accepting "anticipate discharge tomorrow," does the four-fact pass against the source. Allergies: confirmed, nothing added or dropped. Medications: reconciled, intact. Active problems: pneumonia, present and correct. Abnormal values: she opens the actual overnight chemistry rather than trusting the summary's characterization of it, and there is the potassium of 6.1, sitting in the source, nowhere in the summary. The word "stable" evaporates on contact with the real number. She holds the discharge, orders the ECG, and treats the hyperkalemia. The catch did not come from reading the summary more skeptically. It came from refusing to let the summary's word "stable" substitute for the source's number, and going to look.

This is the practical shape of verifying against the source, and it has a rule embedded in it worth naming: a summary's characterization of a fact is not the fact. "Labs reassuring," "vitals stable," "no acute distress," "exam unremarkable" are all summaries of underlying data, and each is a place where a real abnormality can be compressed into a comforting phrase. When a summary offers you a characterization in one of the four harm-carrying classes, that is exactly the moment to go to the source and see the underlying value, because the characterization is where omission hides. You are not verifying against your memory of the patient or against the summary's own internal consistency. You are verifying against the record, the one artifact that contains what actually happened, and you are doing it specifically at the points where a soothing generality has replaced a specific number.

Verifying against the source also settles a question clinicians ask about fabrication: how do you know a stated value is real and not invented. The same way. The summary says "troponin negative," you glance at the troponin in the source. The summary lists metformin, you confirm metformin is on the actual medication list. This is not paranoia and it is not re-charting the whole patient; it is a targeted confirmation of the handful of facts you are about to act on, checked against the place they either exist or do not. The discipline is identical in both directions. For fabrication you ask "is this stated thing actually in the record," and for omission you ask "is everything in the record's four classes reflected here," and both questions are answered in the same place: the source, not the summary, and never the AI's own assurance that it summarized faithfully. Verify, do not repeat blindly, is not a slogan. It is the instruction to point your eyes at the record before your pen touches the order.

Building and Keeping the Habit

The habit only works if it fires every time, and the reason is now familiar: omission gives no warning. Fabrication sometimes announces itself with an implausible claim; omission never does, so you cannot run the check "when something looks wrong," because a dangerous omission looks like nothing at all. The check has to be unconditional, tied to the act of relying on a summary for a decision, not to any feeling that a problem exists. A conditional check is no defense against a failure mode whose defining feature is that it triggers no condition.

Keep the check fast so it survives a busy shift. Four fact classes, confirmed against the source, is a matter of glancing at the allergy list, the med list, the problem list, and the relevant results, and asking two questions of each: is everything the summary claims actually here, and is everything here actually in the summary. On most patients this is under a minute and confirms what you expected. Its entire value is the rare patient where it does not, where the aspirin allergy was invented or the penicillin allergy was dropped, and your one-minute check just prevented a harm that careful reading would have missed entirely. That is the trade: a small, boring, unconditional cost against a rare, catastrophic, invisible one. It is the same bargain as every safety check in medicine, and it earns its place on exactly the same terms.

One last habit makes the whole thing durable: when the cross-check does catch something, close the loop in the record. If the summary fabricated an allergy, do not merely delete it from your mental model; make sure the working record you act from is correct, so the fabrication does not resurface in the next handoff or the next summary. If the summary dropped a real allergy, confirm that the source allergy field is intact and that your orders and communications reflect it. The goal is not just to protect this decision but to keep the error from propagating past you. A caught error that you quietly work around still lives in the record to catch the next clinician; a caught error you correct is genuinely defused. Catching both fabrication and omission is the skill; making sure your correction sticks is what turns that skill into safety that outlasts your shift.

Key Takeaways

• Every summary fails in exactly two directions: fabrication (adding what is not there) and omission (hiding what is there). Naming both precisely is the foundation of catching them.
• The two are asymmetric: fabrication is catchable by reading the summary because the false fact is present on the page; omission is not, because it leaves no trace to read.
• Omission cannot be caught by reading the summary harder. Attention operates on what is present, and an omission is defined by what is absent. The only cure is leaving the summary and checking the source.
• The instinct to "read more carefully" is half right and half dangerous: it helps against fabrication, does nothing against omission, and produces false confidence that you have checked.
• The defense is a fast, bidirectional cross-check against the source: summary-to-source confirms nothing was fabricated, source-to-summary confirms nothing was dropped.
• Spot-check four fact classes every time, because these carry the harm: abnormal values, active problems, medications, and allergies.
• The check must be unconditional, tied to the act of relying on the summary, not to any feeling that something is wrong, because a dangerous omission looks like nothing at all.
• Reading alone catches the harmless fabrication and misses the dangerous omission, the exact inversion of what patient safety requires. Only the source-directed check corrects that.