When an AI Chart Summary Leaves Something Out
AI summaries help clinicians read long records quickly. An omitted allergy or anticoagulant is hard to spot, and the summary itself may not be saved.
Consider a hypothetical. An older patient is admitted overnight with a hip fracture. The admitting clinician has hundreds of pages of prior records and uses an AI tool that summarizes the patient’s history. The summary lists diabetes, heart failure, and kidney disease. It does not mention that a cardiologist started a blood thinner three weeks earlier, because that visit was documented in a separate outside record. The admitting orders add a second anticoagulant for clot prevention, and the patient bleeds.
AI summarization is one of the most appealing uses of AI in hospitals, because the record has grown too large for any clinician to read in full. A good summary saves real time. But a summary is a selection, and the most dangerous error a summary can make is the one that leaves nothing visibly wrong on the screen.
Why omissions are hard to catch. When a summary states something false, a clinician who knows the patient may notice. When it leaves something out, there is nothing to notice. The reviewer would have to know what should have been there. That makes omission the error most likely to pass through human review, and it is why the safety of a summarization tool depends heavily on the safeguards around it.
A recognized category of harm. If harm results, an event like this may fall within the National Quality Forum’s Serious Reportable Events list, which includes “Patient harm associated with a medication error.” An event qualifies only if it is serious and largely preventable as the list defines those terms, and the classification is a reporting and improvement framework, not a legal finding. Medication errors are also commonly categorized by where in the process they began and how far they travelled toward the patient; the NCC MERP Taxonomy of Medication Errors is the common reference for that kind of analysis. In this hypothetical, the error began in the summary, was carried into the admission orders, and reached the patient.
The first question: was the summary saved? Some AI summaries are stored in the chart. Others are generated on screen when a clinician asks and are not kept as part of the medical record. If the summary was not saved, it may be impossible to show exactly what the clinician saw, and the case turns on indirect evidence. This is worth establishing early, along with whether the vendor keeps any log of what was generated.
What evidence may exist.
- The summary itself, if stored, and any vendor-side record of what was generated and when.
- What the tool could see. Which records and sources it drew from, and whether outside records, such as the cardiology visit in the hypothetical, were available to it at all.
- Whether the clinician went beyond the summary. The audit trail can show which source documents and medication lists the clinician opened before ordering.
- Medication reconciliation records. Admission reconciliation is a separate safeguard designed to catch exactly this kind of gap. Who did it, when, and from what sources?
- Clinical decision support. Did a duplicate-anticoagulant alert fire, and how was it handled?
- The tool’s version, configuration, approval, training materials, and monitoring reports.
What the guidance expects. The Joint Commission and the Coalition for Health AI (CHAI) 2025 guidance on responsible AI use is voluntary. It recommends checking that AI tools rely on up-to-date data, monitoring their output quality against known information, training staff on each tool’s limits, and treating AI-related safety events like other patient safety events. For a summarization tool, the limits that matter most are which sources it can see and which it cannot.
What to expect from the defense. A summary is an aid to reading the record, not a replacement for the safeguards hospitals already use: medication reconciliation, pharmacist verification, and drug interaction alerts. Expect the defense to show that those safeguards existed and were reasonable, and that the outside record was not available to anyone. The plaintiff evidence that answers it is the audit trail showing the clinician relied on the summary alone when the medication history was one click away, or governance records showing the organization deployed the tool without telling clinicians what it could not see. Either way, an omission by an AI tool does not by itself establish negligence. The evidence shows where the information was, who could have seen it, and which safeguards had the chance to catch the gap.
Sources
- Joint Commission and National Quality Forum (NQF). Aligning Patient Safety Event Reporting: 2025 Updates to Sentinel Events and Serious Reportable Events. Washington, DC: NQF, January 2026 (Updated May 2026). jointcommission.org. Content reproduced with credit to Joint Commission and National Quality Forum.
- Joint Commission and Coalition for Health AI. The Responsible Use of AI in Healthcare (RUAIH). 2025. jointcommission.org
- National Coordinating Council for Medication Error Reporting and Prevention. Taxonomy of Medication Errors. nccmerp.org
The scenario above is hypothetical. Disclosure: Dr. Bansal is Chief Medical Officer of Diagnose Data, Inc., a healthcare technology company. This article does not describe or endorse any product.
This article is educational and does not constitute legal advice. Safety classifications and voluntary guidance describe reporting and improvement frameworks, not a legal standard of care or any outcome. Sources are listed at the end of the article.