A fellow on Reddit recently described watching a junior resident paste most of a de-identified note into a chatbot and type, more or less, “what should I do next?” The answer came back fast and looked reasonable. That is the problem worth thinking about, and it is one of the quieter decisions in our guide to surviving residency: which parts of your thinking you hand to software during the only years built to grow that thinking.
Residents seem to be splitting into two camps. One group uses AI for nearly everything. The other refuses even the ambient scribe their program is pushing, on the theory that every task handed over makes physicians easier to replace. Both camps are arguing about the tool. The more useful question is about the trainee.
What the deskilling evidence shows so far
The clearest data point comes from gastroenterology, not from chatbots. In a study published in Lancet Gastroenterology & Hepatology in August 2025, endoscopists at four centers in Poland began using an AI tool that flags polyps. When researchers then looked at the colonoscopies those same endoscopists did without the tool, the adenoma detection rate had fallen from 28.4% to 22.4% over roughly three months. It was a retrospective analysis of 1,443 procedures, and the authors framed the drop as a possible effect rather than a proven one. Still, these were trained specialists losing ground in a skill they already had.
Residents face a sharper version of that risk, because many of the skills aren’t there yet to lose. Medical educators have started using three terms for this, which trace back to a 2025 NEJM review on supervising AI use in training: deskilling (losing what you had), never-skilling (never building it because the machine did it first), and mis-skilling (absorbing the tool’s errors as if they were knowledge). Of the three, never-skilling is the one an intern should worry about most. It doesn’t feel like anything. Your notes get done, your plans look complete, and nobody, including you, can see the reps that didn’t happen.
Practicing physicians see it too. In the AMA’s 2026 survey of nearly 1,700 physicians, 81% reported using AI professionally, and 70% said they were very or somewhat concerned about skill loss among current medical students and residents. The attendings evaluating you are using these tools daily and doubting that you should lean on them the same way.
Sorting lookups from judgment
The line that holds up is between retrieving information and deciding what it means for this patient. One resident’s complaint captures the retrieval side well: the two-minute questions, like a dosing protocol buried in a forty-page institutional PDF, eat more of the day than the hard ones. Handing those to a tool costs you almost nothing in training value. Neither does cleaning up a discharge summary you already reasoned through.
The other side is the work an attending will eventually expect to happen in your head at 3 a.m. with nothing open: building the differential before you see anyone else’s, committing to a plan and defending it, deciding which abnormal value matters and which can wait. A workable rule is to do those steps first, on paper or out loud, and only then check yourself against the tool. If the chatbot’s list has something yours didn’t, that gap is the lesson. If you ask it first, you get the answer without the lesson.
It helps to ask one question before opening any AI tool: would I be embarrassed if my attending asked me to walk through this without it? If yes, do the walk-through first.
Where the privacy rules come in
The tool itself matters too. Typing protected health information into a public chatbot that has no business associate agreement with your hospital is treated as an unauthorized disclosure under HIPAA, according to analysis drawn from two JAMA viewpoints. De-identifying first reduces that risk without erasing it, since details can still point back to a person. The practical step is to ask your program or GME office which tools are approved, use those, and keep identifiers out of everything else. This is educational information, not legal advice; your institution’s compliance office is the one that sets the rules you’ll be held to.
Why the radiology fear misreads the tools
The career anxiety tends to land on radiology, while most of the new tools are language models aimed at notes, patient messages and differentials, which touches nearly every specialty. Looking at the AMA numbers, the leading use among physicians is summarizing research and standards of care, at 39%, with assistive diagnosis at 17%. The tools are being folded into physicians’ work, not run in place of them.
That shapes what an employer will want from you in a few years. Expect to be asked which AI tools you’ve used and how. A strong answer names the tool, what you used it for, and how you checked its output. When you interview, ask your own questions back: which tools the group requires, who reviews AI-drafted notes before they’re signed, and whether productivity targets changed when the tools arrived.
The physician who stays hard to replace is the one who can tell when the tool is wrong. You only get that skill by doing the reasoning yourself enough times during training to recognize a bad answer on sight. Hand off the lookups freely. Guard the reps that build your judgment, because residency is the last time anyone will schedule them for you.