In today’s hospitals, the fastest-growing category of AI is not the kind that reads imaging scans or makes diagnoses—but a far quieter one: the AI medical scribe. It listens in on doctor–patient conversations and automatically generates draft medical records; doctors then edit and sign them. According to data cited by researchers from Carnegie Mellon University and the University of Pittsburgh School of Medicine, one integrated health system used this tool to document over 2.5 million patient encounters in a single year—making it one of the most rapidly adopted generative AI applications in healthcare. In other words, the next medical record you receive may not have been written by your doctor—but reviewed by them.
Here’s the question: Has anyone taught them how to review AI-generated medical records?
## The Time Saved Is Real—But What’s Being Measured Isn’t Education
A recent paper published in npj Digital Medicine brings this question squarely into view. First, it acknowledges the real benefits: One real-world evaluation cited in the paper—spanning five academic medical centers, tracking over two years, and involving 8,581 clinicians—found that physicians using the AI scribe saved an average of 13 minutes per 8-hour shift in electronic health record (EHR) documentation time; residents saved 94 minutes—about seven times the average. The paper also cites a small Yale pilot study: After two months of use by 47 trainees, cognitive load scores dropped from 56.3 to 43.3, with adoption reaching 91.5%.
Yet the authors point out a critical blind spot behind these numbers: To date, what has been measured are documentation time, cognitive load, and burnout—metrics valuable for assessing whether the healthcare workforce can sustain itself, but offering almost no insight into what this shift means for *medical education*. More subtly, early evidence suggests the benefits accrue disproportionately to residents and trainees—the very group whose clinical competence is meant to be forged through writing medical records. Those who save the most time are precisely those who should be gaining expertise *by* writing records.
## Writing Medical Records Was Never Just Clerical Work
Why is shifting “from authoring to editing” an educational issue? One of the paper’s authors is himself a medical student—and his statement at the time of publication was direct: He belongs to the “first generation of students who may edit AI-generated records as often as they write their own,” while medical education lags behind this transformation. Reviewing AI-generated documentation may appear to be a simple editing task, yet it entails synthesis, prioritization, and judgment—skills historically honed over years of independent record-writing.
In other words, writing medical records has never been merely about producing a document—it is a training ground for clinical reasoning. As trainees organize a record, they are forced to ask: Which symptom matters most? Which pieces of information conflict? Is the diagnostic evidence sufficient? Now that AI has taken over this training ground, the opportunity to practice clinical reasoning vanishes—and no course fills that gap. The paper concludes: The competencies required to review AI-generated records are specific and teachable—not natural extensions of clinical practice—and current curricula contain nothing directly targeting them.
## Four New Foundational Competencies
The most actionable part of this paper breaks “knowing how to review AI-generated medical records” into four concrete, teachable, and assessable competencies:
**First, error detection.** Can the clinician identify factual errors, omissions, misattributions, and hallucinations—instances where AI confidently states information that simply does not exist? **Second, clinical alignment.** Even if all facts in the record are correct, the record itself may still be flawed: Does it reflect the authentic clinical reasoning of *this* encounter—or is it merely a string of accurate phrases arranged without clinical coherence? **Third, independent generation.** When AI is unavailable, inappropriate, or declined by the patient, clinicians must still be able to produce reliable, high-quality records unassisted. **Fourth, trust calibration.** Dependence on AI output should vary across contexts—emergency vs. outpatient settings, familiar vs. unfamiliar conditions—knowing when to trust more and when to question more is itself a skill.
Each competency comes with defined learning objectives, simulation-based exercises, and assessment methods. The authors emphasize that these competencies anchor to a paradigm far more stable than any specific product: *AI generates; humans review.* So long as this paradigm holds, these four foundational competencies remain relevant—regardless of which underlying model is deployed.
## Intentionally Embedding Errors in Simulation
How do we teach these skills? The paper proposes simulation-based instruction—and assigns it a dual role. As a *teaching space*, simulation centers can systematically embed errors typical of AI scribes into standardized patient encounters, allowing trainees to practice error detection in a zero-risk environment, followed by targeted feedback from instructors on missed items—a feat nearly impossible in real clinical practice. Debriefing after simulation can leverage existing supervisory structures: Educators probe learners’ reasoning processes, assess their use of evidence, deliver individualized feedback, and tailor instruction to identified gaps. As a *research space*, the same infrastructure can generate evidence that real-world clinical deployment never yields: Do trainees with specialized training truly detect more errors in flawed AI-generated records? Can trainees raised alongside AI still independently write sound records when AI is absent? At what stage of training should review competencies be introduced—without displacing foundational authoring skills?
None of these three questions currently has an answer. The authors themselves acknowledge this framework is preliminary—the four competencies, their sequencing in curricula, and their assessment methods all await empirical validation. Yet the paper ends with a weighty statement: The value of medical education lies in its willingness to recognize the emergence of a new competency *before*, rather than *after*, harm occurs.
## Three Takeaway Messages
If you’re a patient, consider asking a new question: Was this record generated with AI assistance—and was the reviewer trained to assess it? If you’re a healthcare system leader, remember that adoption rates and satisfaction scores are not endpoints—saving 94 minutes is real, but if that time isn’t converted into deliberate training, what you’ve purchased is convenience, not capability. If you care about medical education, watch closely those three unanswered questions—especially the one about whether trainees can still write records independently *without* AI. That question will tell us earlier than any framework whether this generation of physicians is becoming stronger—or more dependent.
Finally, a question worth debating: Since writing medical records has long served as a vehicle for training clinical reasoning, should medical students be *required* to practice writing records under “no-AI” conditions—just as pilots today still train in manual flight? Or is that merely a ritual of hardship masquerading as pedagogy—one that should yield to more efficient, newly essential competencies?
Source: Raka, Van Deusen & Wiberg, “From authoring to editing,” npj Digital Medicine (2026)
https://doi.org/10.1038/s41746-026-03351-5