At the end of 2021, four endoscopy centres in Poland brought in AI tools for finding adenomas, small growths in the bowel that can turn into cancer. From then on, whether a colonoscopy used the AI depended on the date of the examination. Nineteen experienced endoscopists, each with more than 2,000 colonoscopies behind them, now worked some days with the software and some days without it.
Researchers later went back over the examinations done without it. In the three months before the AI arrived, these doctors had found at least one adenoma in 28.4% of patients.
In the three months after, on the days they worked without it, they found one in 22.4%.
What the hands forget
The study’s title has a word for what that drop, a fifth of their previous rate, may be: deskilling, a person getting worse at a skill because a tool now does the part that used to keep it in practice. With the AI switched on, the same centres found adenomas in 25.3% of patients.
The authors were careful. The study was observational, so other things may have changed during those months, and all the doctors were experienced, so the effect on newer ones is unknown. A commentary in the same journal still called it “the first real-world clinical evidence for the phenomenon of deskilling.”
The idea is much older than AI. In 1983 the engineering psychologist Lisanne Bainbridge wrote that “physical skills deteriorate when they are not used,” and drew the uncomfortable conclusion for anyone who supervises a machine: “a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one.” Her paper on the ironies of automation is about what happens when that person has to take over.
Pilots heard it first
Aviation got there first. In January 2013 the US Federal Aviation Administration told airlines that “continuous use of those systems does not reinforce a pilot’s knowledge and skills in manual flight operations,” and could degrade “the pilot’s ability to quickly recover the aircraft from an undesired state.” It asked them to make sure pilots had chances to fly by hand, such as “during low workload conditions.”
Six months later, Asiana Flight 214 hit the seawall at San Francisco. The investigators found that the airline’s policy “emphasized the full use of all automation and did not encourage manual flight,” and that the pilot flying, with 9,684 hours, “lacked critical manual flying skills.” Without more chances to fly by hand, the NTSB wrote, pilots’“airplane handling skills degrade.”
Skills that never form
The colonoscopy study found a drop that fits the idea of experienced doctors losing some skill. The NTSB found an experienced pilot lacking a critical one. AI adds a second worry: newcomers who never build the skill at all.
In early 2026 Anthropic ran a randomized trial with 52 software engineers, each of whom had used Python at least weekly for over a year, learning Trio, a Python library none of them had used. One group had an AI assistant in the sidebar that could produce the correct code at any time. They finished about two minutes faster, a difference too small to be statistically significant. On a 27-point quiz about the concepts they had just used, they averaged 4.15 points lower than those who coded by hand. The widest gap was on debugging, the skill a person needs to tell when code, including AI-written code, is wrong.
Afterward, the researchers sorted the AI group by how people had used the assistant. This was an observation, not a second experiment, but the patterns were striking. Those who leaned on it to write or debug the code averaged under 40%. Those who used it to understand, asking for explanations alongside the code or asking only conceptual questions and writing the code themselves, averaged 65% or more. The authors note the sample was small and the quiz came right after the task.
The reconciliation nobody can do
Imagine a finance team that buys software to do the month-end bank reconciliation. It works well for a year. The analyst who used to tie out the accounts by hand moves to another job, and her replacement has only ever approved the software’s summary. Then the bank changes its file format, and the matching starts quietly going wrong. The summary still looks tidy, and nobody left on the team knows what a correct reconciliation looks like from the inside.
Keep the software. Then decide which skills you will need on the day it fails, usually the ones that check its work, and keep them in use on purpose. Have people do a share of the work by hand when the stakes are low, the way the FAA wanted pilots to fly in quiet stretches. For people still learning, set the expectation that AI is for understanding: ask it why, have it explain, try the debugging yourself first. And now and then, measure how people do without the tool.
In Poland, the drop showed up because the trial switched the AI off on some days. A team that only ever looks at results produced with its tools can miss a drop like that. Practice runs, or an occasional test of skills without the tool, can bring it into view.