Issue 013 looked at who is allowed to overrule an AI denial on the payer’s side of the table. This issue moves the question inside your own chart. AI now drafts the note, and as of this year it does so at scale. A multisite study in JAMA found AI scribes already in routine use across major health systems. A peer-reviewed review found that the same tools produce omissions and occasional invented findings. And the World Health Organization drew the line that ties the two together. The machine can write the draft. A named human still owns what it says, and that ownership lives at the moment of review.
Signal 1
A Multisite JAMA Study Found AI Scribes Already in Routine Use Across Five Health Systems, Cutting Documentation Time but Used Consistently by Only a Third of Adopters (April 1, 2026)
Published in JAMA on April 1, 2026, a study by Rotenstein, Holmgren, Thombley, and colleagues tracked 8,581 ambulatory clinicians across Mass General Brigham, Emory Healthcare, UC San Francisco, Yale New Haven Health, and UC Davis between June 2023 and August 2025, comparing the 1,809 who adopted AI scribes with the 6,772 who did not. Adoption was associated with about 13 fewer minutes of total electronic health record time and 16 fewer minutes of documentation time per eight hours of scheduled care, roughly half an additional visit per week, and an estimated 167 dollars in added monthly billing per clinician. The gains were largest for primary care clinicians, advanced practice clinicians, and high-frequency users. After-hours record work did not fall significantly, suggesting the freed time was redirected rather than recovered. One finding is worth holding onto. Only about 32 percent of adopters used the scribe in half or more of their visits, the threshold tied to the largest benefits. Sources: Rotenstein and colleagues, JAMA, April 1, 2026.
What this means for you
This is the largest controlled evidence to date that AI authorship of the clinical note is no longer experimental. It is ordinary, running quietly across academic systems in millions of visits. That is the part worth sitting with. The question has shifted from whether to let a tool draft your documentation to what your obligations are once it does. The same study shows adoption is uneven and undertrained, which means many of the notes being generated are produced by clinicians still learning where the tool helps and where it drifts. Efficiency is the headline. The standard of care for reviewing what the tool wrote is the part that has not caught up.
Signal 2
A Peer-Reviewed Review of Ambient AI Scribes Found They Cut Documentation Burden, and Also Produce Omissions and Occasional Hallucinated Findings (2026)
A narrative review published in the journal Cardiovascular Diagnosis and Therapy synthesized the evidence on ambient AI scribes, the tools that listen to a visit and draft the clinical note. The benefit is consistent and real. The tools reduce documentation burden, lower cognitive load, and return the clinician’s attention to the patient. The risk is also consistent, and less discussed. The same systems produce frequent omissions and occasional clinically significant hallucinations. Across this body of work, reviewers and other studies have documented notes that recorded physical examination findings, and in some cases entire examinations, that never took place during the encounter. The omissions are in some ways the harder failure to catch, because a finding that was dropped leaves nothing in the note to flag. The implication is not that the tools should be set aside. It is narrower and more demanding. An unreviewed AI note is not yet a medical record. It is a draft that looks like one. Sources: Razaghi and colleagues, Cardiovascular Diagnosis and Therapy, 2026.
What this means for you
The draft a scribe hands you can be fluent, complete-looking, and wrong in ways that do not announce themselves. A fabricated exam reads exactly like a real one, because the same model wrote both. That is what makes the review step load-bearing rather than optional. It is the point in the workflow that converts a plausible draft into an accurate record, and it is the one part of the process the tool cannot do for you, because it requires the memory of what actually happened in the room.
Signal 3
The World Health Organization Published a Discussion Paper Drawing One Line Through AI in Health Work: Augment, Do Not Automate (June 2, 2026)
On June 2, 2026, the World Health Organization published a discussion paper on AI and evidence-informed health policy, developed jointly by its Department of Data, Digital Health, Analytics and AI and its Department of Science for Health. The paper maps where AI enters the work of defining problems, designing options, and judging impact, and it names the risk that appears at each stage, from data bias that skews how a problem is framed to monitoring tools that quietly drift a decision away from its original goal. Its operational guidance is concrete. Before a tool is deployed, run an algorithmic impact assessment and a technology readiness review. Once it is in use, pair automated retrieval with human verification, keep a human decision gateway in the loop, and convene multidisciplinary oversight panels combining domain, methods, and ethics expertise. One principle runs through all of it. AI should augment, not automate. The human stays responsible for framing the question, judging the quality of the evidence, and interpreting the result in context. Sources: World Health Organization, June 2, 2026.
What this means for you
WHO is describing the work of making policy, but the principle does not stop at that door. The same sentence governs the note your scribe tool just drafted. Augment, not automate, means a tool can assemble the record, and you remain the one who judges whether it is true before it stands as yours. The paper is useful precisely because it is not about any one product. It states the standard a practitioner can apply to every AI that produces work in their name. The tool extends your reach. It does not inherit your judgment, and it does not take on your accountability.
The Pattern
Put the three together and the shape is clear. The JAMA evidence shows that AI authorship of the clinical note is now ordinary, not experimental, running across major systems in millions of visits. The accuracy evidence shows that the draft can be fluent and wrong in the same breath, with examinations recorded that never happened. And the World Health Organization drew the governing line, augment and do not automate, with the human responsible for judging the result. The question for 2026 is not whether AI should write your documentation. It already does. The question is narrower. Who is the author of a record that a machine produced. The answer taking shape, from a global health body to the journals measuring the tools, is that authorship belongs to the person who reads and signs, not to the system that drafted. The efficiency is real. So is the responsibility, and it has not moved off you.
One Thing You Can Do This Week
Pick one AI-drafted note from this week and read it the way a reviewer would, not the way an author skims their own work. Check the exam findings against what you actually did. Look for what is missing, not only for what is wrong. If you find a fabricated or an absent finding, you have learned something specific about where your tool fails, and that is worth writing down. If the note is clean, you have practiced the one step that, by every account this year, turns an AI draft into a record you can stand behind.