AI discharge summary generation uses large language models to draft the hospital course, diagnoses, medication changes and follow-up plans from a patient’s record. Studies find the drafts are about as good as physicians’ overall, often more concise, but they contain more errors, including omissions and invented details. They can save time only if a clinician checks them properly.
A discharge summary is one of the few documents in medicine written for someone who was not there. The hospital team knows what happened. The primary care doctor, the community nurse, the specialist who sees the patient in six weeks, and the patient, do not. They will rely on a page or two to know what was found, what changed and what still needs doing.
That is why discharge summaries are an obvious target for generative AI and a dangerous one. They are slow to write, often late and frequently incomplete, which makes the case for automation strong. But they are also handoff documents: their whole purpose is to be trusted by a reader who cannot easily check them. A hallucinated detail in a clinical note may be caught by the team that wrote it. A hallucinated detail in a discharge summary travels, unchallenged, to people who have no way of knowing it is wrong.
The central argument of this piece is simple. Among all the places hospitals are putting generative AI, discharge summaries and handoffs are where hallucination risk and patient-safety risk overlap most. That does not mean hospitals should not use AI here. It means the review step is the product, not an afterthought.
Why are discharge summaries such a problem?
The weaknesses of discharge communication are old and well documented. A systematic review in JAMA in 2007, led by Sunil Kripalani, found that discharge summaries were available at the first post-discharge visit in only 12% to 34% of cases, and still missing in many cases four weeks later. Summaries often lacked key information: diagnostic test results were missing from 33% to 63%, and tests pending at discharge were missing from 65%. Direct communication between hospital and primary care physicians was rare.
Nearly two decades later, electronic records have made summaries more available, but not necessarily more useful. Many are long, copied forward from daily notes, and structured for billing and documentation requirements rather than for the reader who has to act on them. Physicians write them at the end of busy shifts, sometimes days after discharge.
Handoffs within the hospital have the same fragility. The I-PASS study, published in the New England Journal of Medicine in 2014, introduced a structured handoff program for residents across nine hospitals. Medical errors fell by 23% and preventable adverse events by 30%, without lengthening handoffs. The lesson was that the structure and completeness of handoff information has measurable safety effects. That is exactly the property generative AI could improve, or quietly degrade.
How good are AI-generated discharge summaries?
The best evidence so far is encouraging with an important caveat.
A cross-sectional study published in JAMA Internal Medicine in 2025, from the University of California, San Francisco, compared hospital course narratives written by physicians with those generated by a large language model for 100 inpatient hospital medicine encounters. Twenty-two attending physicians reviewed them blind. The two were comparable in overall quality and reviewers had no clear preference. The AI narratives were more concise and more coherent but less comprehensive. They also contained more unique errors per summary, an average of 2.91 compared with 1.82 for physicians, spanning inaccuracies, omissions and hallucinations. The potential for harm was low for both, though somewhat higher for the AI versions, and one AI narrative, against no physician narratives, reached the threshold for potential permanent harm.
A second study, published in JAMA Network Open in August 2025 from NYU Langone Health, tested hospital course summaries generated by a model built into the electronic health record. Residents needed to edit a smaller share of the AI summaries than of physicians’ summaries, and attending physicians judged the AI versions more complete. But the AI summaries also contained more confabulations.
An earlier NYU study, in JAMA Network Open in 2024, used an AI model to turn discharge summaries into patient-friendly language. Readability improved dramatically, from about an 11th-grade to a 6th-grade level, and understandability scores jumped. But physician reviewers raised safety concerns in 18 of 100 reviews, mostly about omissions, with some inaccurate statements. The authors concluded that initial use would require physician review.
The pattern across these studies is consistent. AI drafts are fluent, well organised and often as good as or better than what busy physicians write. They also make a different kind of error, and make it more often.
Why are AI errors in handoffs especially dangerous?
Three features make handoff documents a worst case for hallucination.
First, the reader cannot easily verify. A primary care physician reading that a patient was “started on apixaban for newly diagnosed atrial fibrillation” has no reason to doubt it, and may not have access to the full inpatient record. If the model invented the diagnosis, or merged two patients’ details, the error propagates into the next prescription.
Second, omissions are silent. A missing pending test result, a missing instruction to stop a medication or a missing follow-up appointment does not look like an error. It looks like nothing. The UCSF study found AI narratives were less comprehensive, and the Kripalani review showed that pending results were already the most commonly omitted item. AI that is concise at the cost of completeness targets exactly the weakest point.
Third, fluent text invites trust. A well-written summary looks more authoritative than a messy one. The physician reviewing an AI draft at the end of a shift is more likely to skim a polished paragraph than to check each statement against the record. We have made this argument before about AI-drafted patient messages: the real risk is the doctor who agrees with them. Discharge summaries raise the stakes, because the reader downstream is trusting that review happened.
A worked example: the medication that wasn’t stopped
Consider a hypothetical 78-year-old admitted with a fall and found to have low blood pressure. The team stops one of her two blood pressure medications and plans to recheck her kidney function in a week because a blood test was borderline. A urine culture is still pending at discharge.
An AI model drafts the discharge summary from the notes. The hospital course is clear and well written. It correctly lists the admitting problem and the workup. But it picks up the medication list from an earlier note, before the change, so both blood pressure drugs appear as continued. It summarises the kidney function as “stable”. It does not mention the pending culture, because no note explicitly flagged it as pending.
The discharging physician, busy with three other discharges, reads the narrative, finds it accurate as far as it goes and signs. The primary care practice restarts both medications on the basis of the summary. No one rechecks the kidney function. The culture result, which shows an infection, sits in the hospital system until someone happens to notice.
This scenario is illustrative, not drawn from a real case. But each error in it corresponds to a documented weakness: outdated medication reconciliation, omission of pending results, and fluent text that invites a skim. None of them requires the model to invent something dramatic. Ordinary incompleteness is enough.
Is the risk worse than today’s handwritten summaries?
The strongest counterargument is that physician-written summaries are already bad, and an honest comparison has to start there. The UCSF study found physician narratives had errors too, and potential for harm was low for both. Summaries written by exhausted residents at midnight, copied forward from inaccurate notes, days after discharge, are not a safety standard. If AI produces a good draft instantly, the physician can spend the saved time checking the parts that matter most, and the summary reaches the next clinician on the day of discharge instead of the following week. Faster, clearer and slightly different errors may be a net gain.
That argument is serious and partly right. Timeliness alone has safety value, and the NYU work suggests AI drafts can need less editing than human ones. But the argument depends on the saved time actually being spent on verification. The evidence from other AI documentation tools, and from human nature, suggests that time savings tend to be absorbed by more work, not better checking, a dynamic we explored in Ambient AI Scribes Will Reduce Burnout Only If Nobody Spends the Time They Save. The comparison that matters is not AI versus human authorship. It is reviewed AI versus unreviewed AI, and most workflows are not built to guarantee the first.
What this means for hospitals, clinicians and vendors
For hospital leaders, treat AI discharge summaries as a patient safety intervention, not only a productivity one. That means structured review, not a signature. Require the signing clinician to confirm specific high-risk elements: medication changes with reasons, pending results, follow-up actions and who is responsible for each. Build these as checkboxes or structured fields that the model populates and the clinician verifies, rather than hoping they are in the prose.
Audit a sample of AI summaries regularly against the record, focusing on omissions, and share the findings with clinicians. Track downstream measures such as medication discrepancies at follow-up, and whether pending results were acted on.
For clinicians, read AI drafts as you would a medical student’s: likely well written, possibly wrong, and missing things they did not know mattered. Check the medication list against the final orders every time.
For vendors, make verification easy. Link each statement to its source in the record, highlight anything that changed during the admission and flag items the model is less certain about. A summary tool that makes checking faster is more valuable than one that makes writing faster.
For patients and families, ask for a written list of medication changes and pending tests before leaving hospital, and bring the discharge paperwork to your first follow-up appointment. If something in it does not match what you were told, ask your clinician.
Frequently asked questions
Can AI write hospital discharge summaries?
Yes. Large language models can draft discharge summaries and hospital course narratives from the medical record, and several health systems are testing or using them. Studies show the drafts can be comparable in quality to physicians’ summaries and often more concise. They also contain errors, including omissions and invented details, so a clinician must review and correct them before signing.
How accurate are AI-generated discharge summaries?
In a 2025 JAMA Internal Medicine study of 100 hospital encounters, AI narratives were comparable in overall quality to physicians’ but contained more errors on average, 2.91 per summary versus 1.82, and were less comprehensive. Potential for harm was low for both. Accuracy depends on the model, the quality of the underlying notes and the review process.
What are the risks of AI in clinical handoffs?
The main risks are hallucinated details, omissions of important items such as pending test results or medication changes, and over-trust in fluent text. Because readers of handoff documents usually cannot check the original record easily, errors can pass to the next clinician and into treatment decisions. Structured review of high-risk elements reduces these risks.
Can AI make discharge instructions easier for patients to understand?
Evidence suggests it can. A 2024 NYU Langone study in JAMA Network Open found that AI-transformed discharge summaries were far more readable, dropping from about an 11th-grade to a 6th-grade reading level. However, reviewers raised safety concerns in 18 of 100 reviews, mostly about omissions, so physician review remains necessary before patients receive them.
Who is responsible for errors in an AI-drafted discharge summary?
The clinician who signs the summary remains professionally responsible for its content, and the hospital is responsible for the process and tools it deploys. AI drafts do not change that. This is why hospitals should design review workflows that make verification of key elements explicit rather than relying on a signature at the bottom of a fluent document.
Generative AI can make the discharge summary faster, clearer and on time for the first follow-up visit. It can also make it wrong in ways that look exactly like right. The difference is whether someone checks.
If your hospital uses AI to draft discharge summaries, what does the reviewing clinician actually check? We’d like to hear how you’ve built it.
