Keep your hand in

When endoscopy centers in Poland began using an AI that flags polyps, their experienced endoscopists got worse at finding precancerous growths on their own. In the three months after the software arrived, the share of patients in whom they found an adenoma during colonoscopies done without it fell from 28 percent to 22 percent, compared with the three months before. The study was observational and cannot prove that the machine caused the decline. It is still the plainest warning yet of a problem aviation met long ago: a skill handed over to a machine begins to fade.

Most of the argument about AI and doctors concerns which jobs it will take. For a practicing physician the more useful question is which skills to keep, and how. The skills worth having are the ones machines cannot supply, together with the competence to catch a machine when it is wrong. All of them decay without use, and AI is quietly taking away much of the everyday use that kept them sharp. Doctors should practice them on purpose, and employers should give them the time.

The line between what machines can and cannot do in medicine runs through the hands. Anthropic's economic index finds that AI covers few of a radiologist's listed tasks "because AI can't do the hands-on or administrative work in their job profiles," even though it does well at the reading and reporting that fill the working day. A company that has logged a million virtual hospital medicine visits says its remote hospitalists handle the full range of inpatient duties except hands-on procedures such as central lines and intubation, and robots that intubate are still mostly in testing. On the screen side of the line the machines are formidable: in a trial published in March, an AI working alone was right on almost nine in ten written cases, against three in four for physicians' own answers using conventional resources.

What the machines leave to doctors comes down to four skills. The first is the hands: the examination, the procedure, the ultrasound probe. The second is presence. Evaluators preferred a chatbot's written answers to medical questions posted online over doctors' 79 percent of the time and rated them empathetic far more often, so the durable skill lies in the room: sitting with a frightened patient, delivering bad news and being trusted with the decision that follows. The third is judgment that someone must answer for. The federation of America's state medical boards holds that "the physician is ultimately responsible for the use of AI," and that failing to apply human judgment to its output breaches a doctor's professional duties. The fourth is the one most often forgotten: the ability to tell when the machine is wrong.

That last skill depends on a doctor's own competence. In a multicenter study published in September, junior clinicians reviewing GPT-4o's output on simulated cases caught fewer than one in six of its hallucinations; the authors noted that the "clinician-in-the-loop" safeguard assumes "that clinicians can reliably identify and correct these hallucinations." When researchers planted wrong suggestions in software presented as AI, inexperienced radiologists reading mammograms chose the right category less than one time in five, against about four times in five when the suggestion was right, and even veterans were misled, though less often. Doctors can check a machine only with skills of their own, and trainees who lean on one from the start risk what educators have begun to call never-skilling: not acquiring those skills at all.

Skills like these fade without use. The American Heart Association's guidelines note that CPR skills "often show decay by as early as 3 months" after training. Volume seems to matter in the same way: in a large observational study of elderly patients in American hospitals, those treated by older hospitalists were more likely to die within 30 days, except when the older doctor saw a high volume of patients. Nor can training be relied on to supply the skills in the first place: in a multi-site survey of internal-medicine residents, only 30 percent felt confident doing bedside procedures. AI caused none of this, but it adds to it, because every task a machine takes over is practice a doctor no longer gets.

The strongest objection is that this is nostalgia. Nobody asks an accountant to do long division, and if machines are right more often than doctors, practicing what they do better looks like a waste of scarce time. The objection fails on two counts. No machine is close to taking over the examination or the procedure. And for the rest, machines fail in the moments that matter most, on the rare case, the confident error and the system that goes down, which is exactly when a doctor must take over and when a doctor out of practice is least able to. Pilots were warned of exactly this in 2013, when America's air-safety regulator told airlines that "continuous use of autoflight systems could lead to degradation of the pilot's ability to quickly recover the aircraft from an undesired state," and encouraged them to promote manual flying when appropriate.

Building the skills takes practice with feedback, and the evidence that practice pays is good. When residents trained on a simulator before placing central lines in patients, bloodstream infections in their intensive-care unit fell by more than four-fifths. Skill is also visible to others: when peers rated videos of surgeons at work, those judged least skilled had more complications and higher mortality, which is an argument for being watched and coached. Even the hard conversation can be taught: in a randomized trial, oncology clinicians given training, a structured guide and reminders for serious-illness conversations documented them for more patients, earlier, and more often recorded what mattered to the patient.

Keeping the skills is mostly a matter of dose and timing. The heart association recommends following a course with brief booster sessions, weekly or monthly, rather than relying on the course alone, and cites a trial in which nurses given more frequent boosters had better CPR skills a year later. The same logic should apply well beyond resuscitation. Doctors should set a yearly floor for each procedure they mean to keep, log every case and teach the skill to others. They should take regular stretches without AI for the reads and diagnoses they want to own, as pilots fly by hand. And they should commit to their own assessment before looking at a model's: in the March trial, physicians who did so were about as accurate as those shown the AI's answer first, and had done the thinking themselves.

Employers have the bigger job, and many are heading the wrong way: a fifth of physicians told Doximity they already face higher expectations for productivity because of AI. Hospitals should spend some of the time AI saves on practice instead, through protected hours, simulation, booster sessions and procedures shared among doctors rather than routed to a single service. Medical staffs at accredited hospitals renew privileges every two or three years and must review each doctor's performance data at least once a year in between; both reviews should ask for evidence that doctors still do what their privileges allow. Residency programs should have trainees commit to a diagnosis before they consult a model, and test them without one.

The Polish endoscopists had each done more than 2,000 colonoscopies before the study began. A few months of letting software share the looking seems to have been enough to blunt what those years had built. Every doctor who now works beside a capable machine faces the same slow drift, and the remedy is the one pilots were given: keep your hand in, deliberately and often, so that the skill is still there on the day the machine is wrong.