Agentic AI — S2 · 6/10: Remedies for deskilling
While deskilling is widely reported in the media, I am left with one question: is there a remedy for it?
A study published in 2025 in The Lancet Gastroenterology & Hepatology compared what experienced endoscopists could still see without assistance, three months before and three months after an AI tool was introduced. Their detection rate fell from 28.4% to 22.4%. Seniority did not protect them. The decline appeared during normal use of the tool.
In this example, the effect of AI on the user’s skills was studied only after the fact. As with any medical device, the trials that preceded deployment had measured the benefit of the tool, the performance of the human-AI diagnostic system, not what a user lost through continuous use.
This is not specific to one company or one sector. The phenomenon was described as early as 1983, in the context of industrial automation.
The knowledge sat in one field, the decisions in another. After a high-speed deployment of AI, the time for prevention has already passed.
So I looked for an industry that automated long before us, to see what remedy it had found.
For more than a decade, aviation has tried to preserve its most critical skills in the face of automation. The obligations in force since 2019, in Europe and in the United States alike, mainly cover rare, high-risk situations. I read this as a trade-off: the obligation stops where the cost begins.
These provisions do not seek to restore a lost skill. They seek to maintain it. And maintaining presupposes that there is still something to maintain.
Then I looked for what can be recovered.
In a randomised trial published in 2025 in PNAS Nexus, blocking mobile internet on phones for two weeks improved sustained attention. But what improves there is a cognitive function, not a professional skill coming back. Vigilance that slackens and skill that erodes arise from different mechanisms, even though those mechanisms are treated together.
The authors observe that technology often outpaces our ability to understand its consequences, whether for AI or for a mature technology such as the smartphone.
In searching for a remedy, I found a great deal of work on maintenance. On the return of a skill eroded under AI assistance, none. Some may exist.
One can always retort: “retraining will be enough.”
I have found nothing that says whether it works in this context, how long it takes, or at what cost. Too many unknowns to put forward a realistic timeframe.
I was hoping for a remedy. I did not find one.
Have you come across a documented case of a professional skill, diminished by prolonged technological assistance, that was restored?