There is a new piece of workplace research making the rounds, built on interviews with AI experts and executives, and it lands on an uncomfortable idea. Companies using AI heavily might look more capable than ever on paper, while the actual skills of the people doing the work quietly fall apart underneath.
The researchers call this a capability mirage. Work looks finished and polished because AI smooths out the rough edges, but nobody can tell anymore whether the person who produced it actually understands the subject or just typed a good prompt.
This is not a hunch. A study published this year in The Lancet Gastroenterology and Hepatology tracked doctors performing colonoscopies with AI assistance that helps spot early signs of cancer. After the doctors used the AI regularly for three months, researchers turned it off and measured their unassisted performance. Detection rates fell from about 28 percent to about 22 percent. The tool that was supposed to make them better had made their own eye for the job worse.
The same pattern shows up far outside medicine. Pilots who rely on autopilot for long stretches lose manual flying sharpness in as little as two months without practice, a finding well known in aviation safety circles long before generative AI existed. The lesson is not new: skills that go unused fade, and AI is spreading that effect into office jobs at a much faster pace.
Trust is the second casualty, and the numbers here are striking. A global survey by KPMG covering nearly 50,000 workers across 47 countries found that more than half admitted to hiding their AI use and presenting AI-generated work as their own. A separate study from IT firm Ivanti put the figure at around one in three. People are not necessarily trying to cheat; many are just trying to keep up with heavier workloads. But the side effect is that managers have lost a reliable way to know who on their team can actually do the work without help.
For a business leader, the practical risk is not that AI produces bad work today. It is that nobody finds out how much real capability is left until a moment when the AI gets something wrong and there is no backup plan, no senior person who still remembers how to do it the slow way.
The fix does not require rejecting AI tools. It requires treating AI output the way you would treat a new hire's first drafts: check it, ask who is accountable for it, and keep giving people chances to do the work unaided often enough that the skill does not quietly disappear. Companies that skip this step are not saving time. They are borrowing it, with interest due later.