The junior analyst has never built the model by hand. The AI does it, and the AI is usually right, so the skill of doing it, and of knowing when the answer smells wrong, never forms. For a while nothing bad happens. Then a case arrives that the tool gets subtly wrong, and there is no one in the room who can tell.
This is the deskilling risk, and it is the quiet cost that runs underneath AI's productivity gains. Automation that removes drudgery is a gift. Automation that removes the practice through which judgment is built is a loan, and the bill comes due when the tool fails and the human who was supposed to catch it has lost the ability.
Atrophy hides behind good results
The danger is easy to miss because, in the short run, the results look fine. The BCG trial that measured AI's benefits also exposed the trap. On a task past the model's real competence, consultants using AI did worse than those without it, because they accepted a confident wrong answer. The researchers described this as a tendency to stop steering, to fall asleep at the wheel. That failure is not only a momentary lapse; repeated over months, the habit of not steering becomes an inability to steer. You cannot catch an error in a skill you no longer practice.
Deskilling and over-trust are the same disease at different tempos. Over-trust is the acute version, shipping one plausible-but-wrong answer. Deskilling is the chronic version, the slow erosion of the underlying competence that would have let you notice. The more capable the tool, the more seductive it is to hand over not just the labor but the thinking, and the thinking is exactly the part that decays when unused.
Workers sense what they want to keep
People are not naive about this. A large audit of what workers actually want from AI found a consistent preference: automate the low-value tasks, but keep shared agency on the high-value ones, where the AI executes and the human retains judgment and responsibility. Read as a statement about deskilling, that preference is self-protective and wise. People intuit that outsourcing judgment on the work that defines their expertise is how they become unable to do that work at all.
Learning is something you do
As Ricardo Semler has told it, the Lumiar schools he helped start ran for two decades without rows, bells, or memorized answers, organizing learning instead around a project, a tutor, and a deadline, on the conviction that children learn by doing the thing, not by receiving it. The same is true of professional judgment. It is built by wrestling with the problem, getting it wrong, and correcting, not by being handed the finished answer. An organization that lets AI hand its people finished answers for every task is, without meaning to, dismantling the apprenticeship through which judgment used to form.
Keeping the muscle
Guarding against deskilling does not mean refusing the tool. It means being deliberate about which capacities must not be allowed to atrophy, and keeping them exercised. In practice that looks like a few habits. Have people periodically do the hard task without the tool, so the underlying skill stays live. Make verification a real act of judgment rather than a rubber stamp, so checking the AI keeps the checking muscle strong. Protect the apprenticeship: let less experienced people build the thing sometimes, even when the AI could do it faster, because the point is to grow a person, not just finish a task.
The risk is sharpest for the people entering the workforce now, who may never build the underlying skill in the first place. A veteran who leans on AI is spending down a competence already earned; a newcomer who leans on it from day one may never accumulate it. That is why protecting the apprenticeship is not nostalgia but succession planning: the senior judgment an organization relies on to catch the model's errors is grown, slowly, from junior people doing hard things themselves, and a pipeline that skips that step produces experts who cannot actually check the tool they depend on.
The organizations that thrive with AI over the long run will be the ones that used it to remove drudgery while jealously guarding judgment, treating the ability to know when the machine is wrong as the most valuable skill they have, and the one most easily lost. The tool can do the work. It cannot, yet, notice when it has done it badly. That job is still human, and only stays human if humans keep practicing it.
Sources
- Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Lakhani et al. (2025). Navigating the Jagged Technological Frontier (Organization Science). RCT with 758 BCG consultants: on tasks inside AI's frontier, GPT-4 users did 12.2% more tasks, 25.1% faster, ~40% higher quality; on tasks outside it they erred more. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Shao, Zope, Jiang, Pei, Brynjolfsson, Yang et al. (2025). Future of Work with AI Agents (arXiv 2506.06576). Across 1,500 workers, 844 tasks and 104 occupations, people want AI to automate low-value tasks but prefer shared agency (AI executes, human keeps judgment) on high-value work. https://arxiv.org/abs/2506.06576
- Chan et al. (2024). Visibility into AI Agents (ACM FAccT '24). As work is delegated to autonomous agents, accountability requires agent IDs, real-time monitoring and activity logs, each with trade-offs for privacy and concentration of power. https://doi.org/10.1145/3630106.3658948