Endoscopists who spent years performing colonoscopies with AI assistance missed 6 percentage points more precancerous polyps on days when the AI system was unavailable. Not because of worse training, because the trained eye had learned to outsource pattern recognition to a machine. That finding, originally published in The Lancet Gastroenterology and Hepatology, was one of the first in a growing stack of evidence that Nature bundled in June 2026 into a clear conclusion: the longer professionals work with AI, the worse they perform when the AI is gone. Four practical agreements prevent that from happening in your team.
What does the research actually show?
The evidence for AI-induced deskilling is now replicated across multiple fields and multiple study designs.
The strongest signal comes from the ACCEPT trial in gastroenterology, run across four endoscopy centers. Experienced physicians worked alongside an AI system that flagged suspicious polyps during colonoscopy procedures. Before the system was introduced, the physicians detected adenomas, precancerous growths, in 28.4% of procedures. After months of AI-assisted work, detection rates on days without the system dropped to 22.4%. Researcher Yuichi Mori of the University of Oslo concluded that clinicians became "less motivated, less focused, and less responsible" when making cognitive decisions without AI.
“There is no established solution against deskilling right now. It should be a very hot research topic in the next decade.”
Yuichi Mori, University of Oslo, in Nature (June 2026)
A separate Anthropic study, a randomized controlled trial with 52 software engineers, found that using AI assistance led to a statistically significant 17% reduction in skill mastery, equivalent to nearly two letter grades. Participants who used AI to produce code rather than to understand it scored below 40% on a post-task knowledge assessment. Those who used AI for conceptual questions and explanations scored 65% or higher. The pattern is identical to what happens in medicine: outsourcing cognitive tasks to AI erodes the judgment that makes oversight possible.
A May 2026 global survey by GoTo, covering 2,500 workers across industries, confirms the breadth of the problem. 39% of employees believe AI has weakened their skills. Among Generation Z that rises to 46%. Thirty percent, nearly one in three, say they can no longer function normally without AI assistance. A Wolters Kluwer survey of US healthcare workers found 77% of physicians and 70% of nurses worried about skill erosion through AI use.
Why does this hit your organization harder than a hospital?
Hospitals have structural safeguards against skill erosion. There is peer review, complication tracking, colleagues who watch each other's work.
In a business team of ten to fifteen people, almost none of those safeguards exist.
Here's the thing: it is like navigating by GPS every day for a year, then having to find your way through a city you have visited a dozen times without a map. You knew the route. But the skill of navigating independently has quietly worn away from disuse.
According to TheAIDaily's Shadow AI SMB Research, 44% of SMB employees share company data with free AI tools, an average of 3.1 times per week. Most of those organizations have no policy governing when AI should and should not be used. Fewer still have any plan for keeping human skills current.
The skills that disappear first
Routine tasks don't fade through AI use. Formatting a spreadsheet or scheduling a meeting remains just as easy after a year of AI use.
What erodes quietly are the skills that require judgment.
Evaluative judgment. The ability to tell whether something is right without re-checking it. The endoscopists lost exactly this: the eye that spots a suspicious polyp without a red border drawn around it by a machine.
Process knowledge. Understanding how something works, not just what the output should be. If your team generates reports with AI for months, within half a year no one recalls which data belongs in the report, which sources are reliable, or which exceptions matter.
Quality calibration. Recognizing what "good" looks like. The GoTo survey found that 43% of workers have already used AI-generated content of low quality in their work, not out of laziness, but because the internal compass for quality has slowly drifted.
Anomaly detection. The instinct that something is off. An experienced accountant who "notices something strange" in a balance sheet before the numbers confirm it. A project manager who senses a schedule is wrong before running the math. This may be the most dangerous skill to lose, because it is also the hardest to train deliberately.
Five warning signs your team is too dependent on AI
Watch for these patterns:
- Nobody reviews AI output anymore. The standard response to an AI-generated report is "looks good" with no substantive check.
- New hires learn the role through AI. They know how to write a prompt, but not how the underlying process works.
- When the AI goes down, work stops. During an outage or usage limit, the team waits for the tool rather than switching to manual methods.
- Errors surface later than they used to. Clients, suppliers, or auditors catch mistakes the team would once have spotted themselves.
- More trust in AI than in human judgment. 28% of workers in the GoTo study trust AI more than themselves on certain tasks. That is the point at which human oversight has effectively been switched off.
Recognize three or more of these? That is not a signal to purchase more AI licenses. It is a signal to implement the four agreements below.
Four agreements you can make this week
Banning AI would be counterproductive. But you can build deliberate moments of unassisted work into your workflow so that core skills don't silently erode. These four agreements cost little time and produce a team that uses AI as an amplifier rather than a replacement.
1. A monthly AI-free practice day. Pick one day a month when your team completes a critical process entirely without AI, not every process, only the ones where errors carry serious consequences. An accounting firm might manually close one month-end per quarter. A marketing team might develop one client strategy without AI assistance. The day doesn't need to be conventionally productive. It exists to prove that your team still holds the skill.
2. Rotate the AI reviewer. Make sure the person reviewing AI output is not the same person who wrote the prompt. Rotate this weekly. Fresh eyes catch more errors, and the reviewer gets practice using judgment at the same time. For context: the GoTo survey found 83% of workers worried about liability for AI errors. Rotating review keeps skills sharp and distributes accountability.
3. Onboarding without the AI shortcut. Let new employees complete a core process manually for the first two weeks before getting access to AI tools. A junior accountant who has manually filed a VAT return understands what the AI is doing when it later takes over. A junior marketer who has written a campaign brief by hand recognizes when AI output is below standard.
4. The quarterly stress test. Each quarter, ask one simple question per department: "Could this team run this process if AI were unavailable for a week?" If the answer is no, you have found a vulnerability. Address it before the AI actually goes down, not after.
What tasks can AI simply take over?
For any task you hand to AI, ask yourself: if the AI makes a mistake here, would someone on your team catch it?
- Yes, immediately. Scheduling, meeting notes, standard emails, data entry. Let AI handle these freely. An error in meeting notes gets caught by the first reader.
- Maybe, with delay. Financial analysis, client reports, legal summaries. Rotating review (agreement 2) is essential here.
- No, no one would notice. Strategic assessments, risk analysis, quality control. These are exactly the processes where your team should practice without AI most often.
The less visible an AI error is, the more important it is that your team retains the ability to do it without help.
What does this have to do with the EU AI Act?
The EU AI Act, in force since February 2025, requires organizations to ensure AI literacy among employees working with AI systems. That obligation goes well beyond knowing that AI tools exist. It means employees must be able to critically evaluate AI output and understand the limitations of the systems they use.
Worth noting: if AI dependency erodes your team's ability to assess AI output for accuracy, your organization may not be meeting that requirement in practice. The EU AI Office and national data protection authorities across member states are increasingly focused on exactly this kind of compliance gap. The four agreements above are also a way to make AI literacy demonstrably real, not just documented on paper.
What do you do with this tomorrow?
Start with the quarterly stress test. This week, walk through each department and ask the question: "Could we run our core work if AI went down for a week?" Write down where the answer is no. Those are your vulnerabilities.
Then schedule the first AI-free practice day for the process where dependency is highest. Not as punishment, as training. The way a pilot lands manually on a regular basis even when the autopilot would handle it perfectly.
The Anthropic research on coding skills found that how people use AI matters as much as whether they use it at all. Engineers who asked conceptual questions scored 65% or higher. Those who just delegated code generation scored below 40%. The same principle applies to any knowledge role. When deciding whether to hire or automate, keep in mind that effective AI delegation requires the human side to remain capable enough to evaluate what AI produces. The four agreements in this article are how you keep that side intact.
