A new primary care physician inherits a patient with a fifteen-year chart: hundreds of notes, three prior practices, a decade of labs. She asks the AI to summarize the record, and it returns a clean paragraph opening with "72-year-old man with a longstanding history of seizure disorder on levetiracetam." She builds her whole first visit around that anchor. Except the patient has never had a seizure. A single note in 2014 recorded "rule out seizure," a resident carried it forward as "seizure disorder," it copied itself through a hundred subsequent notes, and now the AI, faithfully summarizing what the record says, has laundered a fifteen-year-old error into a confident, front-and-center diagnosis. The summary is not hallucinating. It is accurately reporting a chart that has been wrong for a decade.

The Longitudinal Record Is a Different Beast

Summarizing a single encounter is one problem. Summarizing years of accumulated notes is a categorically harder one, and the difference matters for how you verify. A longitudinal summary is asked to pull a coherent narrative out of a record that was never written to be coherent. It was written one visit at a time, by many hands, across different systems, with different concerns, over years. The story the AI constructs is a genuine synthesis, and synthesis is powerful: it can surface a pattern no single note contains, a slow trend, a recurring complaint, a medication history that only makes sense assembled. When it works, a longitudinal summary can make you smarter about a patient in ninety seconds than an hour of scrolling would.

But that same synthetic power is exactly what makes the longitudinal summary dangerous in a specific way the single-encounter summary is not. To build a coherent story, the model must decide what the record "means," and it builds that meaning out of whatever the record says, including everything the record got wrong. A single-encounter summary can drop today's abnormal value. A longitudinal summary can do that and inherit a decade of accumulated error, elevate a long-propagated mistake into the headline diagnosis, and present the whole thing with the smooth confidence of a story that finally makes sense. The risks are not just larger; they are different in kind, and they have names.

It is worth being explicit about why this categorical difference matters for your verification habit. In the previous lesson, the danger was mostly omission: the single-encounter summary dropping today's abnormal value, a failure you counter by confirming plan-changing facts at the source. That habit still applies here. But the longitudinal summary layers three additional dangers on top of it, and none of them is caught by simply checking whether a fact is present in the record, because in every case the fact is present in the record. It is just wrong, or over-corroborated, or drawn from an incomplete view. You cannot catch these by asking "is this in the chart," because the answer is always yes. You catch them only by asking a harder question about where the fact came from and whether the record ever earned the confidence it now projects.

Distinguish, as you go, between two failure modes that feel similar and are governed by completely different defenses. Omission is when the summary leaves something out: the potassium of 6.1 from this morning that never made the paragraph, the falling hemoglobin trend that got smoothed into "stable anemia." Fabrication is when the summary invents something the record never contained: a diagnosis no note supports, a medication no one prescribed. The single-encounter lesson trained you to guard against both by confirming plan-changing facts at the source. The longitudinal summary adds a third category that is neither, and that is why it is so easy to miss: the summary neither omits nor fabricates. It faithfully reports something that is present in the record, repeated across the record, and wrong. Your omission radar does not fire, because nothing is missing. Your fabrication radar does not fire, because the summary invented nothing. The fact is there, it is corroborated, and it is false, and only a provenance question reaches it. Knowing which of the three you are hunting tells you which question to ask, and asking the wrong question leaves the most dangerous error untouched.

Inherited Error: The Chart Was Already Wrong

The first and most important risk of longitudinal summarization is one the AI does not create and cannot detect: the record it is summarizing was already wrong. Charts accumulate errors. A tentative diagnosis becomes a definite one. A medication the patient stopped taking three years ago still lives on the active list. A "penicillin allergy" that was actually a childhood stomach upset hardens into a documented contraindication. These errors were there before any AI touched the chart, and a summarizer, doing its job faithfully, will report them as fact, because to the model they are the facts. The record is its ground truth. It has no independent access to the patient to know the record is lying.

This is a crucial and counterintuitive point about what an AI summary can and cannot protect you from. A summary can be perfectly faithful to the source and still be clinically false, because the source itself is false. Faithfulness to the chart is not the same as truth about the patient, and the AI can only ever offer you the former. When the summary says "seizure disorder on levetiracetam," it is telling you the truth about what the chart contains. Whether that is the truth about the patient is a question the summary cannot answer and does not know it cannot answer. That gap, between what the record says and what is real, is the space where inherited error does its damage, and no amount of model improvement closes it, because the error is upstream of the model.

The clinical shapes this takes are worth cataloguing, because you will meet all of them. A tentative diagnosis hardens into a definite one: "rule out seizure" becomes "seizure disorder," "possible COPD" becomes "COPD," "query PE in 2019" becomes a documented pulmonary embolism the patient never had, carrying with it years of unnecessary anticoagulation. A discontinued medication lingers on the active list because someone stopped it verbally and no one reconciled it, so the summary confidently lists a drug the patient has not taken in three years. An allergy that was never really an allergy calcifies into a hard contraindication: a childhood stomach upset after amoxicillin becomes "penicillin allergy," and now the patient is denied first-line antibiotics for the rest of their life on the strength of a label no one ever re-examined. A social-history note entered once in a moment of frustration ("noncompliant") follows the patient through every subsequent encounter, quietly shaping how each new clinician reads them. None of these are AI errors. Every one of them predates the summarizer and would be in the chart if you scrolled through it by hand. The AI's contribution is not to create them but to gather them, polish them, and present them as the confident headline of a coherent story, which is precisely what makes them harder to doubt than they were when they sat scattered across a hundred messy notes.

A summary can be perfectly faithful to the chart and still be false about the patient. Faithfulness to the record is not truth, when the record is already wrong.

Copy-Forward Propagation: How One Error Becomes a Hundred

Inherited error has a specific engine, and every clinician knows it intimately: copy-forward. The copy-and-paste and note-cloning features that make documentation bearable also make error immortal. A finding entered once, correctly or not, gets carried into the next note, and the next, and the next, until a single keystroke from years ago appears in a hundred notes and takes on the weight of overwhelming corroboration. The record does not distinguish between a fact confirmed a hundred times and a fact copied a hundred times. They look identical on the page: a diagnosis that appears in note after note after note.

This is where longitudinal summarization becomes genuinely treacherous, because a summarizer is, in effect, a corroboration-counting machine. A claim that appears in a hundred notes looks, to a model building a narrative, like a very well-established fact. The more times an error was copied forward, the more confidently the summary will assert it, because repetition in the source reads as strength of evidence. So the copy-forward mechanism does not just preserve the original error; it amplifies the summary's confidence in it, in exact proportion to how many times it was propagated. The fifteen-year-old "seizure disorder," copied into a hundred notes, does not appear in the summary as a hesitant maybe. It appears as the confident opening clause, precisely because it was wrong so many times.

This inverts an instinct many clinicians carry, that a diagnosis appearing consistently throughout a chart is a diagnosis you can rely on. In a hand-written, independently-authored record, that instinct was reasonable: consistency across many separate clinical judgments was real corroboration. In a copy-forward record, consistency is often the opposite of independent judgment. It is a single judgment, made once, mechanically duplicated. The chart that looks most internally consistent may simply be the chart where one early error was copied most diligently. So when a longitudinal summary presents a diagnosis with the smooth authority of "this has been true for years, it appears everywhere," the correct clinical response is not reassurance. It is the specific, trained suspicion that you may be looking at a well-preserved mistake rather than a well-established fact.

Context-Window Limits: The Model Cannot See What It Was Not Given

There is a third, quieter risk that is easy to forget precisely because it is invisible: the model can only summarize what it was actually given. A fifteen-year chart may exceed what fits in the model's context window, the amount of text it can consider at once. When that happens, something upstream, the vendor's pipeline, a retrieval step, a truncation rule, decides which parts of the record the model sees and which it never does. The model then produces a confident, coherent summary of the fraction it received, with no awareness of, and no flag for, the years or the notes that were left out. The summary does not say "based on the 40% of the chart I could fit." It just summarizes, seamlessly, as if it had seen everything.

This means a longitudinal summary can be wrong not because it inherited an error and not because it fabricated, but because a critical note simply was not in the window. The one visit where the patient's true history was correctly documented, the oncology note that reframes everything, the specialist's letter that corrected the propagated error, any of these can be silently outside what the model saw. And you have no way to know, from the summary alone, what was excluded. The confident coherence of the output gives you no signal about the completeness of the input. A summary of half the chart reads exactly as authoritative as a summary of all of it. This is why "the AI read the whole chart" is an assumption you cannot safely make, and why the source, once again, is the only arbiter.

Vendors differ enormously in how they handle this, and most of the difference is invisible to you at the point of care. Some pipelines retrieve the most relevant sections and tell you what they pulled; some quietly truncate the oldest material; some summarize in chunks and stitch the pieces together, which can lose the connections between chunks. You are generally not shown which strategy produced the summary in front of you, and you rarely see a note reading "the following years were not included." This is one of the concrete things the ONC transparency rules and the responsible-AI guidance are pushing toward, the idea that you should be able to ask and be told what an AI intervention actually did. Until that transparency is universal, the safe posture is to treat every longitudinal summary as potentially built from a partial view, and to let that assumption drive you back to the source for the facts that matter most.

Verify Key Facts to the Original Source

The defense against all three risks is the same, and it is the through-line of this entire chapter: verify the key facts against the original source, not against the summary and not against the propagated chart entry, but against the earliest, most authoritative documentation you can find. For a longitudinal summary, this has a particular flavor. When a diagnosis anchors your plan, do not just confirm it appears in the record, of course it appears, that is how it propagated. Trace it back. Find where it entered the chart. Ask whether the original documentation actually supports the diagnosis or whether you are looking at a "rule out" that hardened into a certainty through repetition.

The facts that earn this deeper trace are the load-bearing ones: the anchoring diagnoses that shape your whole approach, the medications you are about to continue or build on, the allergies that will constrain your prescribing, and any history that determines a major decision. You will not trace every fact in a fifteen-year chart; that would defeat the purpose. But the two or three facts on which your plan actually rests deserve a moment of genuine skepticism about their provenance, especially when they arrive in the summary with suspicious confidence. A diagnosis that has been asserted a hundred times and never once questioned is not necessarily well-established. It may simply be well-copied.

Tracing provenance is a concrete skill, not a vague exhortation, so it helps to know what the motions actually are. You are looking for the origin of a claim, which means sorting the record by date and reading forward from the earliest mention rather than backward from the hundredth copy. For a diagnosis, the origin you want is the moment it was made: the workup, the imaging, the specialist's letter, the biopsy result, the moment someone did the work and reached a conclusion. When you find a solid origin, a clear workup with an unambiguous conclusion, you trust the diagnosis and stop; you have earned the right to build on it. When you find a soft origin, a "rule out," a "history of per patient," a "presumed," a differential item that was never resolved, you have found a claim the record never actually earned, and its hundred subsequent copies do not upgrade it. And sometimes you find no origin at all: the diagnosis simply appears one day, fully formed, in a note that inherited it from somewhere you cannot locate, which is itself a warning that the chart is carrying a fact no one can vouch for. The tell that tells you to spend this effort is a mismatch between confidence and evidence. When a claim shapes your care disproportionately to how well you understand its basis, or arrives with more certainty than you can immediately justify, that gap is the signal to trace before you build.

The Provenance Question

The single most useful habit for longitudinal summaries is to ask, of any plan-anchoring fact, a provenance question: where did this come from, and does the original support it? This is a different question from "is this in the chart," and it catches a different class of error. "Is this in the chart" catches fabrication. "Where did this come from" catches inherited error and copy-forward propagation, because it forces you past the wall of repetition back to the moment the fact entered the record, where you can actually judge whether it was ever true. When the provenance is a solid workup with a clear conclusion, trust it and move on. When the provenance is a tentative note, an assumption, or a nothing you cannot even locate, you have found an error the summary confidently laundered into a fact.

There is a useful tell that tells you when to spend the effort. The plan-anchoring facts that most deserve a provenance trace are the ones that arrive with more confidence than their apparent evidence warrants, or that shape your care disproportionately relative to how well you understand their basis. A diagnosis that reorganizes your entire approach to a patient, that you did not make yourself, and that you cannot immediately point to a solid workup for, is exactly the fact to trace before you build on it. You are not being paranoid about the whole chart. You are spending a few minutes of skepticism on the two or three facts that, if wrong, would send the entire visit in the wrong direction. That is a proportionate and defensible allocation of a busy clinician's attention, and it is precisely where the harm concentrates.

A Worked Example: The Diagnosis That Was Never True

Return to the inherited patient and watch both paths. The summary opens "72-year-old man with longstanding seizure disorder on levetiracetam." The physician who treats the summary as truth builds her visit around a seizure patient: she reviews his levetiracetam, considers his "epilepsy" in every subsequent decision, counsels him on driving restrictions, and never questions the anchor. She may keep him on an anticonvulsant he never needed for years more, propagating the error one more time in her own confident note, which the next AI will summarize with even more corroboration behind it. The error has now survived another clinician precisely because the AI presented it so cleanly that it invited no scrutiny.

The physician with the provenance habit reads the same opening clause and treats its very confidence as a reason to check, because a "longstanding seizure disorder" is exactly the kind of plan-anchoring diagnosis that earns a trace. She searches the record for the origin, not the hundredth copy but the first mention, and finds the 2014 "rule out seizure" note with a normal EEG and no documented seizure, ever. The workup was negative. The diagnosis was a placeholder that a copy-forward turned into a decade-long fiction. She now has a completely different and correct picture: a man on an unnecessary medication, carrying a diagnostic label that has probably affected his insurance, his driving, and his care for ten years. She documents the discrepancy clearly, so that her note corrects the record rather than propagating the error, and the next clinician, and the next AI summary, inherits truth instead of the mistake.

The two visits diverged at a single moment: whether the clinician accepted the summary's confident anchor or asked where it came from. Same chart, same summary, same fifteen years of accumulated error. One clinician laundered the mistake forward; the other traced it to its source and broke the chain. That is the entire skill of safe longitudinal summarization in one contrast.

Notice, finally, what the provenance habit protects that goes beyond this one patient. Every time a clinician accepts a propagated error without checking, they do not merely fail to catch it; they re-endorse it, adding one more authoritative-looking note to the pile the next AI will read. Errors in a longitudinal record are not static. They are self-reinforcing, growing more entrenched with each clinician who defers to them, because deference looks identical to confirmation in the finished chart. This is why the discipline is not just about protecting the patient in front of you, though it does that. It is about refusing to be one more link in a chain of copied confidence, and instead being the clinician who traces the fact back, tests it against its origin, and either confirms it honestly or corrects it cleanly. In a record that increasingly gets read and rewritten by machines, that human act of tracing a claim to its source is what keeps the whole longitudinal record anchored to reality rather than to its own accumulated momentum.

Key Takeaways

• Summarizing years of notes is categorically harder than summarizing one encounter, because the model must synthesize a coherent story from a record that was never written to be coherent, and it builds that story out of everything the record got wrong.
• The biggest risk is inherited error: the chart was already wrong before any AI touched it, and a faithful summary reports old errors as fact because the record is its only ground truth.
• A summary can be perfectly faithful to the chart and still be false about the patient. Faithfulness to the record is not truth when the record is already wrong.
• Copy-forward propagation turns one keystroke into a hundred corroborating notes, and a summarizer reads repetition as strength of evidence, so the more times an error was copied, the more confidently the summary asserts it.
• Context-window limits mean the model may summarize only the fraction of a long chart it was actually given, with no flag for what was left out; a summary of half the chart reads exactly as authoritative as a summary of all of it.
• The defense is to verify plan-anchoring facts against the original, earliest source, not the propagated chart entry and not the summary.
• Ask the provenance question of any load-bearing fact: where did this come from, and does the original documentation actually support it? This catches inherited error that "is it in the chart" never will.
• When you find and correct a propagated error, document it clearly so your note breaks the chain, giving the next clinician and the next AI summary truth to inherit instead of the mistake.