The checker has to know the work.

A colonoscopy report asks one question that matters more than the rest: did the doctor find an adenoma, the small growth that can turn into cancer? The share of procedures in which a doctor finds at least one is called the adenoma detection rate, and gastroenterologists track it closely, because a doctor who finds more of them prevents more cancers. At the end of 2021, four endoscopy centres in Poland began using an AI system that watches the video feed and flags possible polyps. As part of a trial, procedures were done with the tool or without it depending on the date. Researchers later looked only at the procedures done without it. In the three months before the tool arrived, the detection rate in those procedures was 28.4 percent. In the three months after, it was 22.4 percent.¹ These were experienced doctors, nineteen of them, each with more than 2,000 colonoscopies behind them. With the tool switched on, their rate was 25.3 percent.²

Those numbers measure two different things. One is the finished work: how well the procedure goes with the tool in the room. The other is the judgment to check it: what the doctor can see unaided, which is also what the doctor brings to the screen when the tool flags something, misses something, or is not there at all. Most deployments of AI measure only the first. The Polish trial measured the second almost by accident, because its design required some procedures to be done the old way.

My argument is that generative AI is moving knowledge work from producing answers to checking them, and that checking depends on the expertise that producing used to keep in practice. When people begin from the machine's answer, the finished work can improve while their ability to judge it declines. The cheapest protection a manager has is the order of the work: an unaided attempt first and the tool second, wherever the stakes justify the time. Here is what the argument does not claim. The evidence is young, much of it drawn from surveys, small samples or a single specialty, and none of it shows permanent harm. Several studies find that AI improves both output and learning when it is used with some structure. The claim is narrower: there are two things to measure, and usually only one of them is on the dashboard.

The job moved from making to checking

In 2025 a team from Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers who use generative AI in their jobs and collected 936 examples of real tasks.⁴ Asked where they still applied critical thinking, the workers described a change in what the thinking was for. Less of it went into gathering information and drafting. More went into verifying what the AI had produced, fitting its response into their own work, and what the authors call task stewardship: overseeing a job rather than performing it. The responsibility did not move with the labor. The person still signs the memo.

The study's sharpest finding concerns confidence. In the authors' words, "higher confidence in GenAI is associated with less critical thinking," while workers who were more confident in their own abilities reported applying more of it.⁴ This is a correlation drawn from self-reports, and the authors say so; their respondents sometimes confused spending less effort on the tool with spending less effort on thought. But the pattern is the one you would predict if verification is effortful and trust lets people skip it.

Checking is not a lighter kind of doing

It is tempting to think of verification as the easy part: the AI does the work and a person glances over it. Consider what checking requires. To notice that a drafted contract leaves out an indemnity, you need to know that contracts like it usually include one. To notice that a forecast's growth rate is implausible, you need a feel for what plausible looks like. To see the flat polyp the software missed, you need the eye that finds flat polyps. Recognizing a wrong answer draws on the same knowledge as producing a right one.

A field experiment at Boston Consulting Group shows what happens when that knowledge is not engaged. Researchers from Harvard Business School and several other universities gave 758 consultants realistic assignments, some with access to AI and some without.⁵ On tasks inside the AI's capabilities, consultants with the tool completed 12.2 percent more tasks, 25.1 percent faster, at more than 40 percent higher quality. On a task designed to fall just outside those capabilities, consultants with the tool were 19 percentage points less likely to reach a correct answer than those without it. The authors called the boundary a jagged frontier, because it is hard to see from the inside. The tool sounded just as sure of itself on both sides of the line. The only protection was a person who knew the work well enough to tell which side they were on.

Skill that is not used decays

None of this would matter if the checking skill maintained itself. It does not, and the reason is ordinary. The psychologists Evan Risko and Sam Gilbert define cognitive offloading as using physical action to reduce the mental demands of a task: writing a shopping list, setting a reminder on a phone, tilting your head to read a sideways label.⁶ Offloading is usually sensible. A memory you do not need is a memory you need not maintain.

The trouble comes when the thing offloaded is the judgment itself. A 2026 review in Computers in Human Behavior Reports by Fakhroddin Noorbehbahani and Kiemute Oyibo maps the research on this path.⁷ It begins with automation bias, the tendency to favor an automated suggestion over one's own judgment. Repeated often enough, that acceptance becomes a habit of offloading, and the habit, over time, can become deskilling: the erosion of an ability that is no longer exercised. The review is careful about what the evidence supports. Most of the studies it maps are short, the long-term effects are largely unmeasured, and the direction of the effect depends on how the tool is used. Several of the studies it includes find that AI used appropriately improves reasoning.

The industrial version of the problem is old. Lisanne Bainbridge described it for control-room operators in 1983, and an earlier essay here, "The exception is the business.", followed it to the hard cases automation leaves behind. What is new is the reach. A plant had a few operators. Generative AI reaches everyone who writes, analyzes or codes.

Reading the colonoscopy numbers carefully

Return to the Polish trial, because its numbers are easy to overread. The study was retrospective and observational: it compared two three-month windows rather than randomly assigning doctors to use or avoid AI.¹ Other things could have changed between the windows, and the authors hedge their conclusion accordingly, saying only that continuous exposure to AI might reduce detection. The fall appeared at every centre, but its size varied widely. Rajiv Bhuta, a gastroenterologist at Temple University who was not involved, called the result a signal while noting that each physician contributed fewer than a hundred procedures on average.² Yuichi Mori of the University of Oslo, one of the authors, has pointed out that the non-AI days inside a trial may not resemble ordinary practice.³

So the study does not prove that AI deskills doctors. What it shows is more useful to a manager. An organization that tracked only the gold number, the detection rate with the tool, would have seen 25.3 percent and moved on. The blue number, the doctors' unaided rate, fell six points and would have stayed invisible. Most AI rollouts are built that way. They measure the output with the tool, because that is what the tool was bought to improve, and they have no instrument at all for the skill without it.

The order of the work

If unaided skill matters and erodes, what can be done short of taking the tools away? The strongest evidence comes from a classroom. Hamsa Bastani of the Wharton School and her colleagues ran a randomized experiment with nearly a thousand high school mathematics students in Turkey.⁸ During practice sessions, one group used a tool that mimicked a standard ChatGPT interface. A second group used a version built as a tutor, which offered hints rather than answers. A control group had only their textbooks and notes. With the plain tool, practice grades rose 48 percent compared with the control group. With the tutor, they rose 127 percent. Then the tools were taken away for an exam. Students who had practiced with the plain tool scored 17 percent below the control group. The tutor's safeguards largely removed that harm.

The same underlying model produced both results. What differed was who did the thinking first. The plain tool let students start from its answer. The tutor made them produce an attempt and then helped them improve it. The practice score, the gold number, made the plain tool look like a triumph. Only the exam, the blue number, showed what it had cost.

A smaller and much shakier study points the same way. Researchers at the MIT Media Lab recorded the brain activity of 54 young adults writing essays over four months, some with ChatGPT, some with a search engine and some with neither.⁹ The ChatGPT group showed the weakest neural engagement, and its members "struggled to accurately quote their own work." In a fourth session, a small group who had been writing unaided switched to the tool, and they showed better recall and more brain activity than the group that had used it from the start. The authors named the pattern cognitive debt: effort saved now and repaid later.

The phrase is memorable, and the study deserves to be held at arm's length. It has not been peer reviewed. Only 18 people took part in the fourth session. A published comment by researchers at the University of Vienna and TU Dresden argues that the sample was too small for the number of comparisons made, and that some results are reported inconsistently.¹⁰ The lead author, Nataliya Kosmyna, has herself warned against reading it as evidence of a "brain on vacation" or of long-term change.¹¹ It illustrates the idea. It does not establish it.

Objections, considered

We offloaded arithmetic to calculators and survived. The best form of this objection is that losing a skill a reliable tool covers is the whole point of the tool. We gave arithmetic to calculators, directions to maps and spelling to spellcheck, and we gained more than we lost. True, when the tool is reliable and the skill is not needed to check it. A calculator does not produce a plausible wrong sum. Generative AI is right often enough to be trusted and wrong often enough to need a checker, which is exactly the combination in which a fading skill does damage. Risko and Gilbert treat offloading as normal and often wise; the question is which judgments must stay with the person.

The evidence is too thin to act on. A survey, a review, a preprint and one observational study cannot carry a management prescription. Agreed, and that is why the prescription is to measure, not to restrict. The classroom experiment is randomized and firm, and it points to a cheap change in design rather than a ban. Measuring unaided skill costs a little time. Discovering its absence during an incident costs more.

An unaided first pass gives back the productivity we paid for. The consulting gains inside the frontier were large, and asking people to work without the tool first surrenders some of them. The answer is to use the first pass where a wrong answer is expensive and the checker must be credible, not on every email. In the Turkish study the tutor delivered a larger practice gain than the plain tool and avoided the exam harm, so the trade may be smaller than it looks.

This is a problem for students, not professionals. People who already know their trade, the argument goes, cannot forget it in three months. But the endoscopists in Poland had thousands of procedures behind them, and the Microsoft survey found its confidence pattern among working professionals. Experience may slow the decline. Nothing here shows that it prevents it.

What a manager can do

Three practices follow. They are recommendations, not findings. First, sequence the work. Where a decision is consequential, ask for the person's own estimate, diagnosis or draft before they see the tool's answer, then use the tool to challenge and improve it. Second, measure the blue number. Schedule some work to be done without the tool and record its quality, as the Polish trial did by design. A skill you never observe is a skill you cannot vouch for. Third, give the verification to people who still do the underlying work often enough to be good at it, and treat that practice as part of the job rather than as a lapse in efficiency.

The answer can get better while the person checking it gets worse.

Four questions for your team

• Which mistakes? Which decisions in your area depend on a person catching the machine's mistake?
• When last? When did that person last do the work without the tool, and how good was it?
• Which first? Does the workflow show them the AI's answer before or after they form their own?
• Who measures? Who tracks the skill, and not only the output?

Go back to the colonoscopy report and its one question: was an adenoma found? An organization that wants a human in the loop needs a second line beside it, one nobody thought to print in 2021: what the doctor finds when the tool is off. A loop is only as strong as the person who could close it alone.