Six weeks into my internship, on a Sunday morning at the VA, I ran my first code.
The patient was POD#5 from a CABG and going home that day. The cardiothoracic attending had rounded early and pulled his pacing wires. I had never met the attending. I had never met the patient. I got called because the man went into afib while getting dressed.
The floor nurses weren’t permitted to push beta-blockers, so I did. What I didn’t know was that the surgeon had used a radial artery graft, uncommon at the time, and that the patient’s calcium channel blocker infusion had been stopped that morning. The ’90s were a different era.
An hour later, the attending and the senior resident walked into the room together. By then I was pumping the man’s heart with my bare hands.
Real physician voices, twice a week
Free, and one click to unsubscribe.
He lived. His daughter was in the room the entire time. The family was grateful to me and considerably less grateful to the attending and the VA. My written account of that hour ran two pages.
The record and the hour
I have relived that hour more times than I can count, most of them under oath in federal court. And every single time, I remembered something I had left out of the record or simply gotten wrong.
That isn’t a confession of carelessness. I knew that note would be scrutinized, and I wrote it carefully. It is a description of what a record is. An hour of sight, sound, smell, fear, and physical effort compressed into two pages by a 28-year-old surgery intern whose adrenaline hadn’t worn off. The note captured what I could reconstruct, in the order I could reconstruct it, filtered through what I understood at the time. I didn’t yet understand the most important fact in the room.
Dr. Michael Peck recently made this point about code reviews on this site: They depend on documentation and memory gathered after one of the most stressful events in medicine, and memory is unreliable after that kind of stress. He is right, and he stops one step short.
Where the records go now
Charts are no longer just for quality review and litigation. They are the raw material for the medical AI we are now told will match or replace physicians. Models learn from clinical notes. They are evaluated against case vignettes distilled from clinical notes. When one of them “outperforms doctors,” it usually means it did better on a written summary of a patient that a physician already assembled, deciding what mattered and what to leave out.
No AI has ever walked into a room and taken a history and physical, let alone run a code. It gets the two pages. It never gets the hour. External pacer not catching. My first unsupervised intubation. The 1,001 things going through my mind while cracking his chest. The wrong internal paddles for the pacer. The 15 hours (8 minutes according to the daughter) I straddled that man’s body, rhythmically squeezing his heart, wondering WTF do I do now.
And the problem is invisible by design. What never made it into the record can’t show up as an error. A model trained on my note would learn from it faithfully, including the parts I got wrong and the silence where the parts I forgot should be. The benchmark it is graded on was built the same way, so everything agrees with everything else. The loop is closed, and the loss sits outside it.
What this means
I am not arguing against AI in medicine. I build AI systems in my spare time. I would welcome an AI scribe in a code, not to write the note but to hand me a timeline I can check my memory against. That is the right job for it: a reminder, not a replacement.
But every claim about AI clinical performance should come with a question: performance on what? If the input was a record a physician created, then the physician did the hardest part of the work before the AI ever saw the case. Collecting the history, performing the exam, and deciding what mattered is the medicine. The rest is reading comprehension.
Twenty-seven years later, I still don’t have a complete account of that Sunday morning. I was there. The record wasn’t.
John C. Ferguson is a cosmetic surgeon.
