Jesse Jacoby
At the end of 2021, four colonoscopy centers in Poland started using an AI system that flags possible polyps on the screen during the procedure. The doctors involved were experienced. Each had performed more than 2,000 colonoscopies. Researchers then looked at how those same doctors did when they worked without the tool. Before the AI arrived, they found adenomas in 28.4 percent of their unassisted procedures. A few months after it became routine, that rate had fallen to 22.4 percent, a relative drop of about 20 percent.1
Nobody decided to make those doctors worse. Nobody noticed it happening either, until someone measured it.
That’s the pattern most organizations are now building into their AI programs without seeing it. Every efficiency gain comes with a second entry that almost never appears in the business case: the gradual loss of the human skill needed when the system is wrong, unavailable, or facing something it hasn’t seen before. The gain gets paid out now. The fragility gets billed later. Organizations that want the first without the second have to decide on purpose which skills stay in practice, and then pay for the practice.
A Skill You Don’t Use Starts to Leave
The colonoscopy study is useful because it’s so clean. Same doctors, same centers, a short window of time, and a measurable outcome that matters to patients. The authors, writing in The Lancet Gastroenterology and Hepatology, called it a deskilling risk, and the finding held for people with thousands of procedures of experience behind them.1 Experience didn’t protect them. Regular reliance on the tool changed how they looked at the screen when the tool wasn’t there.
Knowledge work shows a softer version of the same thing. Researchers from Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers who use generative AI at least once a week and found that confidence in the AI was associated with less critical thinking effort. Workers who were sure of their own ability on a task were more likely to question the output. Workers who were less sure tended to accept it.2 The people with the least slack to lose are the ones most likely to stop checking.
Aviation Already Ran This Experiment
Commercial aviation has lived with high automation for decades, and it offers a preview of where this goes. In January 2013 the Federal Aviation Administration issued a safety alert to airlines about manual flying. It said that continuous use of autoflight systems doesn’t reinforce a pilot’s manual skills and could degrade the pilot’s ability to recover an aircraft from an undesired state.3
Three years later the Department of Transportation’s Office of Inspector General looked at whether anything had changed. It found that the FAA had no method for determining how often pilots actually fly by hand. Only two of the nine carriers the auditors visited analyzed how often their pilots used the autopilot. One carrier directed its pilots to fly with the highest level of automation available. The report also pointed to a 2010 study in which pilots overestimated their own manual flying skills.4
One of the most safety-obsessed industries in the world needed a regulator to remind operators that pilots should occasionally fly the plane. Most companies deploying AI have no regulator, no flight data and no simulator. They have a productivity dashboard.
Why the Loss Never Shows Up on the Dashboard
The gains from automation are immediate and easy to count. Cycle time drops. Output per person goes up. Somebody presents the chart at the next leadership meeting. The loss is gradual, spread across hundreds of people, and invisible until the day the system fails or gets something wrong that a practiced person would have caught.
BCG calls this distributed de-skilling, a collective erosion of human skills that weakens an organization’s intelligence and resilience over time. In its survey of 70 senior executives, half said de-skilling is already happening in their organizations and more than 60 percent expect it to become a material threat within three to five years. Almost 90 percent named overreliance on AI outputs without challenge as a concern. Only one company in ten has an organization-wide strategy for dealing with it, and a third haven’t discussed it at all.5 The skill BCG rated most at risk was judgment and decision making.
This is a different problem from the one facing entry-level roles, where the work that used to train people is disappearing before they get to it. Here the people losing skill are the experienced ones, the same people the organization assumes will step in when something goes wrong.
The Fallback Is a Capability, Not a Line in a Plan
Every AI deployment carries an unwritten assumption: if the tool fails, a person will do the work. Business continuity plans usually name the backup system and the recovery time. They rarely name the human fallback, and almost never check whether that person can still do the job at the standard the business needs.
A 2026 working paper by Simrita Singh, Naireet Ghosh and Tinglong Dai models exactly this choice. When a firm uses AI that outperforms its people when it works but sometimes fails, keeping people engaged in the work lowers output today and preserves the skill the firm will need as a fallback later. The authors argue that designing human and AI work becomes a human capital investment decision.6 Their model also finds that a more reliable system can reduce the incentive to keep people practiced, which means the failures that do happen arrive in front of people who’ve had the least recent practice.
That’s the fragility in plain terms. The better the tool works, the less ready the organization is for the day it doesn’t.
Paying the Dividend Back
None of this argues for slowing AI adoption. It argues for budgeting the skill the adoption consumes. A handful of practices cover most of it.
Write down the skills that have to survive. For every AI-assisted workflow, name what a person still needs to be able to do without the tool, and how fast. Some skills can go. Nobody misses manual formatting. Others, like reading a contract for risk or judging whether a forecast is plausible, are the reason the role exists.
Schedule the reps. Aviation’s answer was to require manually flown approaches, departures and upset recovery in training. The business version is a regular AI-off cycle for the skills on that list. BCG describes a large Indian bank that runs monthly AI-free sessions for exactly this purpose.5
Ask for the person’s answer before the machine’s. Shell builds learning paths that require people to define the problem themselves before they use AI.5 The same habit works for experienced staff. Write down your estimate, your diagnosis or your recommendation first, then compare it with the tool’s.
Measure the unassisted baseline. The Polish study only found the decline because someone measured the doctors working without the tool. Most companies never check. Pick a sample of AI-assisted work each quarter, have it done by hand, and compare the quality.
Count it in reviews and promotions. BCG’s recommendation is to make human skill development visible in performance reviews and promotion criteria.5 If the only thing anyone is scored on is output, output is what people will protect.
Efficiency With a Spare
The efficiency gains from AI are real, and most organizations are right to chase them. But a gain that slowly spends down the skill needed when the system fails isn’t free. It’s borrowed against the organization’s resilience, and the bill arrives at the worst possible moment.
The fix starts with one question added to every AI business case. If this tool went dark for a week, who would do the work, and how good would it be? Then test the answer. Choose one AI-assisted process this quarter, run it by hand for a day, and compare the results. That one day says more about the organization’s real resilience than any adoption metric on the dashboard.
References
- K. Budzyń et al., “Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study,” The Lancet Gastroenterology and Hepatology, 2025, as reported in “AI use in colonoscopy may weaken doctors’ detection skills over time,” News-Medical, August 13, 2025.
- Microsoft Research and Carnegie Mellon University, “The Impact of Generative AI on Critical Thinking,” CHI 2025, as reported in “Some workers already let AI do the thinking for them, Microsoft researchers find,” The Register, February 11, 2025.
- Federal Aviation Administration, “SAFO 13002: Manual Flight Operations,” January 4, 2013.
- U.S. Department of Transportation, Office of Inspector General, “Enhanced FAA Oversight Could Reduce Hazards Associated With Increased Use of Flight Deck Automation,” Report AV-2016-013, January 7, 2016.
- Sagar Goel, David Martin, and Charikleia Kaffe, “When Everyone Uses AI, Companies Risk Losing Critical Skills,” BCG, June 2026.
- Simrita Singh, Naireet Ghosh, and Tinglong Dai, “Managing the Human Fallback: Skill Investment Under Improving AI and Worker Mobility,” arXiv working paper, June 27, 2026.
