A hospitalist is discharging a patient at four in the afternoon, the ninth discharge of a heavy day. She asks the AI to draft the discharge summary from the chart, and thirty seconds later it produces something genuinely impressive: a clean hospital course, a tidy problem list, a reconciled-looking medication list, a follow-up plan. It reads like the work of a rested physician with all afternoon. What it does not say, anywhere, is that a blood culture drawn on hospital day three is still pending, that the cardiology follow-up the team promised was never actually scheduled, and that the patient's warfarin dose was changed on the last day. The summary is not wrong in what it says. It is dangerous in what it silently left out, and discharge is the exact transition where a silent omission becomes a readmission, a missed cancer, or a bleed.

Why Discharge Is the Dangerous Transition

Every point where a patient moves between settings or teams is a point where information can fall on the floor, and the discharge from hospital to home or to a next facility is the highest-stakes handoff in medicine. The receiving clinician was not there for the admission. The patient and family, often overwhelmed and newly home, become responsible for a medication regimen that may have changed substantially, for follow-up appointments they have to keep, and for watching for warning signs no one may have clearly explained. The discharge summary is the document that carries the entire hospitalization forward into that vulnerable gap. When it is complete and accurate, it is the safety net. When it drops something, there is often no other net, because everyone downstream is trusting the summary to be the whole story. The evidence bears this out: care transitions are a well-documented source of preventable harm and avoidable readmission, and information that fails to cross the transition is a leading contributor. An AI that drafts the summary faster does not change that underlying risk. It changes only how quickly a plausible-looking document arrives, which means the human verification step becomes more important, not less.

This is why the specific failure mode of AI at discharge is not the one people instinctively fear. The instinctive fear is fabrication, the AI inventing a fact. Fabrication is real and matters, but at discharge the more probable and more lethal failure is omission: the summary that is entirely truthful and quietly incomplete. A generative model summarizing a long, messy hospital course is under implicit pressure to produce something coherent and readable, and coherence is achieved partly by leaving things out. The pending culture that has no result yet is easy to omit because it is an open loop with nothing tidy to say about it. The follow-up that was discussed but never booked is easy to omit because the chart is ambiguous about whether it happened. The last-day medication change is easy to bury because it is one line in a reconciliation the model treats as settled. Every one of these omissions makes the summary read better, cleaner, more finished, which is exactly why they are so dangerous. The danger at discharge wears the disguise of a good summary.

Consider what the model is actually doing when it drafts. It is predicting the most probable next span of text given the chart it was shown and the shape of the thousands of discharge summaries in its training. A discharge summary, statistically, is a finished-sounding document: it resolves its threads, it closes its loops, it reads like the account of a completed episode of care. So the model reaches, by default, for that finished register. The pending blood culture is a thread that is not finished, and a finished-sounding document has no natural slot for an unfinished thread, so the safest prediction, the one that most resembles the training distribution, is to leave it out or to phrase around it. This is not the model being careless. It is the model being exactly what it is: a fluent producer of text that resembles good text, with no independent concept of clinical consequence and no knowledge of which absent facts would kill your patient. The eloquence is real and the judgment is absent, and at discharge that specific combination is the hazard. A tool that wrote clumsily would at least alarm you. A tool that writes beautifully invites the one thing you cannot afford, which is trust proportional to polish.

There is a second-order effect worth naming, because it is where good clinicians get caught. When you have read ten AI drafts and all ten were substantially correct, you begin, without deciding to, to extend the eleventh the same credit. The very reliability of the tool on the easy majority of cases trains your vigilance downward, so that on the one draft that quietly dropped the abnormal potassium or the pending pathology, you are reading with the relaxed attention the previous ten earned. This is automation bias in its purest form: the tendency to accept an authoritative automated output without the checking you would apply to a human colleague, and it grows precisely because the tool is usually right. The discharge summary is where automation bias meets the highest-stakes handoff in medicine, and the meeting is not a coincidence you can wait out. It is a standing condition of the work, which means the countermeasure has to be standing too.

The Three Things That Must Never Drop

Across the discharge failures that reach patients, three categories recur so reliably that they deserve to be a permanent, named checklist you run against every AI-drafted discharge summary before it goes out. Memorize them, because they are the difference between a summary that protects a patient and one that quietly abandons them.

Pending Results

Cultures, pathology, pending labs, imaging read after discharge, send-out tests that take a week: these are open loops at the moment of discharge, and open loops are exactly what a coherence-seeking summarizer prunes. The catastrophe here is concrete and common: a blood culture or a biopsy comes back positive or abnormal after the patient is home, and because the discharge summary never flagged it as pending, no one is watching for the result, no one owns following it up, and the finding, an early cancer, a bacteremia, a drug-resistant organism, sits unaddressed until it presents again, worse. A discharge summary that does not explicitly list every pending result, name who is responsible for reviewing it, and state how it will reach the patient is not a complete summary. Verify the pending list against the source directly. Do not trust the AI to have carried every open loop forward, because carrying open loops forward is precisely the thing it is worst at.

Follow-Up Appointments

A follow-up that is documented as arranged but was never actually scheduled is worse than no follow-up at all, because it creates false reassurance: everyone believes the safety net exists, so no one builds it. The AI cannot know whether an appointment discussed in a progress note was actually booked; it can only see that follow-up was mentioned, and it will often render that mention as if it were a confirmed plan. Verify that each follow-up in the summary corresponds to a real, scheduled appointment with a date, or is explicitly flagged as still needing to be arranged and by whom. The specialist follow-up for the incidental lung nodule, the anticoagulation clinic visit for the new warfarin, the primary care visit to recheck the kidney function: these are load-bearing, and a summary that lists them without confirming they exist is a summary that lies by implication.

Medication Changes

Discharge is where medication regimens are most volatile and where reconciliation errors most often reach a patient. A dose was titrated, a home drug was held, a new anticoagulant was started, an antibiotic course was defined with a stop date. The AI-drafted medication section can look perfectly reconciled while silently getting one of these wrong: carrying forward a home dose that was changed, dropping the fact that a medication was intentionally stopped, or failing to convey the duration of a course. The patient goes home and takes the old dose, or doubles up, or never stops the antibiotic, or stops the anticoagulant they needed. Every medication change during the stay, especially holds, new starts, dose changes, and defined courses, must be verified against the record and clearly conveyed to the patient in language they can act on, not buried in a list that looks settled.

The high-risk drugs deserve a sharper eye. An anticoagulant that is dropped from the summary or rendered at the wrong dose is a gastrointestinal bleed or a stroke waiting on the calendar. Insulin carried forward at an inpatient dose that no longer fits a patient eating normally at home is a hypoglycemic event. An opioid with no taper or no stop point is a dependence risk and an overdose risk. When the AI hands you a medication section, the drugs that changed and the drugs that are dangerous are the two lists you check against the source first, and where they overlap is where you slow all the way down. The reconciliation that matters is not the list of every pill the patient takes; it is the delta between what they took before and what they take now, and whether the summary states that delta in words the patient can act on. A medication list that is complete and a medication list that communicates the changes are two different documents, and the AI reliably produces the first while implying it has produced the second.

At discharge, the AI's most dangerous output is not a lie. It is a beautiful, truthful, incomplete summary, because a summary that reads as finished is a summary no one thinks to check for what it left out.

The Verification Workflow: Check Against the Source, Not the Summary

The core discipline of the AI-assisted discharge summary is deceptively simple to state and hard to do under load: you verify the summary against the source, never the source against the summary. The distinction matters enormously. If you read the summary and ask does this look reasonable, you will almost always say yes, because a fluent summary looks reasonable by construction, and you will catch fabrications but miss omissions entirely, because you cannot notice the absence of something you were not looking for. The only way to catch omission is to work in the other direction: start from the source and confirm that each critical element made it into the summary.

Concretely, that means a short, structured pass. Pull up the pending results in the chart and confirm each one appears in the summary with an owner and a plan. Pull up the appointments and confirm each follow-up in the summary is real, and each needed follow-up is present. Pull up the medication reconciliation and walk every change from the stay, confirming the summary conveys it correctly. This is not rereading the whole chart; it is a targeted cross-check of the three high-omission categories against the primary record. It takes a few minutes, and those few minutes are the entire reason the summary is safe to send. A discharge summary generated in thirty seconds and verified in five minutes is a triumph of the workflow. A discharge summary generated in thirty seconds and sent in thirty-one is a readmission waiting for a trigger.

There is a reason to build this as a fixed habit rather than a judgment call. Discharge happens at the end of the day, at the end of the service, on the busy afternoon, precisely when your vigilance is lowest and the pressure to move is highest. A verification you perform only when you feel careful is a verification that will fail on exactly the discharges where it matters most. The pending culture gets dropped on the chaotic Friday, not the calm Tuesday. So the cross-check against the source is not something you do if the summary looks off; it is something you do every time, by rule, because the summaries that need it most are the ones that look most finished.

It helps to make the pass small enough that it survives a bad day, because a verification routine that is too heavy is a routine that gets skipped, and a skipped routine protects no one. The following is a compact version you can run in a few minutes against the primary record, not against the draft.

Notice what the table is doing. In every row the source you open is the authoritative system, and the summary is the thing being audited against it, never the reverse. That single directional rule is the whole method. A clinician who internalizes the table stops asking does this summary look complete and starts asking does the summary carry forward what the chart says is open, which is a question that can actually be answered and that lights up the gaps.

What You Still Own, and What You Must Tell the Patient

The AI drafted the summary, but the accountability for the transition is entirely yours, and it extends beyond the document to the patient in front of you. A discharge summary is not only a record for the next clinician; it is the basis for the discharge instructions the patient and family take home. The most beautifully verified summary in the world fails its purpose if the patient walks out not understanding that a culture is pending and someone will call, that they must keep the cardiology appointment, and that their warfarin dose changed. Part of owning the discharge is translating the verified content into plain, actionable language, and confirming the patient actually understood it, the pending result, the appointments, and every medication change.

This is also where disclosure and documentation of the AI's role belong. If the patient-facing discharge instructions were generated with AI, the applicable state rules on disclosure apply, and a licensed clinician's review is what keeps the communication compliant and safe. Several states have moved here. A California rule effective at the start of 2025 requires that generative-AI-produced patient clinical communications carry a prominent disclaimer and instructions to reach a human, while exempting communications a licensed provider has read and reviewed, which is one more reason the review is not optional. A Texas statute effective at the start of 2026 requires disclosure of AI use in diagnosis or treatment to the patient. These specifics are an evolving patchwork and the thresholds and exemptions differ by state, so treat any number or effective date as something to verify against current law rather than to repeat from memory. In the record itself, a brief, accurate note that the summary was AI-drafted and verified against the source strengthens the documentation. None of this transfers your accountability to the tool. The summary carries your name, the transition is your responsibility, and if a pending culture goes unaddressed because it was dropped from a summary you signed, the answer to why did no one follow up on this is, unavoidably, that you sent a summary that did not carry it forward. The tool made the draft fast. Keeping the patient safe across the transition is still the job, and it is still yours.

There is a documentation move that is easy, cheap, and protective, and it is worth making a habit. When you disagree with something the draft implied, or when you add back a pending result the AI dropped, a single sentence in the record noting that the summary was AI-assisted and reconciled against the source before signing does real work. It shows a reviewer, months or years later, that a human took ownership at the gate, that the tool produced a draft rather than a decision, and that the verification the standard of care now expects actually happened. The evolving standard of care cuts both ways with AI: a clinician can be exposed for following a wrong machine output and for ignoring a correct one, and the short note explaining what you confirmed and why is the thread that ties your judgment to the record. It costs a sentence. It buys a defense.

Why Omission Is the Failure You Cannot See

It is worth dwelling on why omission is so much harder to catch than fabrication, because understanding the mechanism is what makes the verification discipline stick. When an AI fabricates, it adds something to the summary, and an added thing is a visible object on the page. If the summary claims a procedure that never happened or a finding that is not in the chart, a careful reader can see the false statement and challenge it. Fabrication announces itself; it is present, and presence can be inspected. Omission is the opposite: it is an absence, and absence is invisible by definition. Nothing on the page tells you that a pending culture was left out, because the whole nature of the failure is that there is nothing on the page. You cannot proofread your way to catching an omission, because proofreading examines what is there, and the omission is precisely what is not.

This is why the human mind is so poorly equipped to catch discharge omissions by reading, and why the fix has to be structural rather than attentional. No amount of careful reading of a summary will reveal a gap, because reading operates on the text and the gap is outside the text. The only reliable detector for an absence is a second, independent source that lists what should be present: the chart's pending-results screen, the scheduling system, the medication reconciliation. You compare that authoritative list against the summary and the gaps light up. This is the deep reason the workflow insists on verifying against the source rather than reviewing the summary. It is not a stylistic preference. It is the only method that can detect the class of error that actually harms patients at discharge, and building it into your routine is the single most valuable thing this lesson can give you.

A Worked Example: The Summary That Reads Perfectly

Return to that four-in-the-afternoon discharge. The patient is a 72-year-old admitted with a fall and found to have a urinary tract infection, treated and improving. During the stay three things happened that a good summary must carry: a blood culture drawn on day three was still pending at discharge, the team started apixaban for newly discovered atrial fibrillation and changed the home lisinopril dose, and a follow-up with cardiology for the new atrial fibrillation was discussed but the appointment was never actually booked. The AI draft is excellent. It narrates the UTI treatment cleanly, lists a reconciled medication list, and states cardiology follow-up under the follow-up section. It does not mention the pending blood culture at all, renders the cardiology follow-up as though it were arranged, and lists apixaban without flagging that it is a new high-risk anticoagulant or that the lisinopril dose changed.

The unverified path. The hospitalist reads the summary, finds it reasonable, and sends it. The patient goes home. Two days later the blood culture flags positive for a bacteremia that no one is watching for, because the summary never said it was pending; it is caught only when the patient re-presents febrile and sicker. The cardiology appointment the summary promised does not exist, so the new atrial fibrillation goes unmanaged. And because the medication change was not clearly conveyed, the patient is confused about the new blood thinner. Three separate near-misses, all from omission, all from a summary that read beautifully.

The verified path. The hospitalist runs the source cross-check. Pending results: she pulls the microbiology screen, finds the pending blood culture, and adds it to the summary with an explicit owner and callback plan. Follow-ups: she checks the scheduling system, finds no cardiology appointment actually exists, and either books it or flags it clearly as still needing to be arranged, with responsibility assigned. Medications: she walks the reconciliation, confirms apixaban is new and flags it as high-risk anticoagulation, and confirms the lisinopril change is conveyed. Then she translates all of it into discharge instructions the patient understands and confirms understanding. The AI saved her the drudgery of composing the hospital course from scratch. The five-minute cross-check against the source is what actually made the discharge safe. Same draft, same patient, same busy afternoon. The workflow was the difference between a safe transition and three preventable harms.

Key Takeaways

• Discharge is the highest-stakes handoff in medicine: the receiving clinician was not there, the patient is newly responsible for a possibly changed regimen, and the summary is often the only safety net carrying the hospitalization forward.
• The dangerous AI failure at discharge is not fabrication but omission: a truthful, coherent summary that reads better precisely because it silently left something out.
• Three categories must never drop, and deserve a permanent checklist: pending results (cultures, pathology, send-outs), follow-up appointments (real and scheduled, not just mentioned), and medication changes (holds, new starts, dose changes, defined courses).
• A follow-up documented as arranged but never actually booked is worse than none, because it creates false reassurance that stops anyone from building the real safety net.
• Verify the summary against the source, never the source against the summary: you cannot notice the absence of something by rereading a fluent summary, so you must start from the chart and confirm each critical element made it in.
• The source cross-check must be a fixed habit, not a judgment call, because discharge happens when vigilance is lowest and the pressure to move is highest, and the summaries that look most finished are the ones that need it most.
• Owning the discharge extends beyond the document to the patient: translate the verified pending results, appointments, and medication changes into plain, actionable language and confirm understanding.
• The AI made the draft fast, but the transition is still your responsibility; a pending culture dropped from a summary you signed is, unavoidably, your unaddressed result. AI assists, the clinician decides, the record proves it.