When AI Builds Skill, and When It Erodes It
As AI takes on more of the work we once did ourselves, there’s growing concern about what happens to the skills we’re no longer practicing. For example, a student who solves more problems with AI may be learning more, or may simply have an easier way to get the answers. Much the same goes for an employee handling customer cases, or a management consultant finishing a business report. The perceived productivity gains can make the tool worth using, but they don’t tell us whether the person has become more capable.
In “When AI Thinks for Us” published in Skeptic, professor Hyoungbin Park argues that the effort involved in thinking helps children (and adults) develop abilities, including the ability to engage in moral reasoning, and that handing work over to AI could weaken them. It’s a reasonable concern. But what we know about how young brains learn isn’t enough to establish that AI causes lasting cognitive or moral harm. For that, we need evidence of how people manage after the AI tool is no longer available to them. Can they apply what they’ve learned to an unfamiliar problem? Can they spot a mistake when AI gives them a convincing answer?
Field studies have begun to address these questions, with mixed results. One found that students performed worse when they later had to work without AI. Another suggested that inexperienced employees retained what they had learned with help from AI. Before drawing broad conclusions about what AI does to human skill, we need to understand why these tools had such different effects.
Learning Math
In a randomized field experiment published in the Proceedings of the National Academy of Sciences, Hamsa Bastani and colleagues worked with nearly 1,000 high-school math students. They compared ordinary classroom practice with two kinds of AI assistance. One of the tools was a relatively unrestricted chatbot that could supply answers. The other was a specially designed tutor that encouraged students to engage with the problem rather than simply provide them with the solution.
During assisted practice, both AI groups looked spectacular. Compared with the control group, scores rose 48 percent with the unrestricted assistant and 127 percent with the guided tutor. If a school had evaluated those tools solely by the quality of completed practice work, it might have concluded that the AI had dramatically improved learning.
Then came an exam. Students who had used the unrestricted assistant scored 17 percent worse than students who had never used AI. Their practice results had been a poor indicator of what they could do independently.
The guided AI tutor, however, did not produce a negative effect. But it also did not produce any positive results: its student users did not demonstrate a statistically significant improvement over the control group on the unaided exam. They performed far better during practice, yet the study did not establish that the AI tutor made their independently-demonstrated mathematical ability better than conventional instruction.
An AI tool can improve performance on average, while making some mistakes more likely to survive human review.
This was a study of specific math problems and a specific generation of AI. It did not measure brain development or what happens after years of use. Extrapolating its findings, such as the 17 percent drop in performance in the unrestricted AI tutor use to all education, would be an error. Nevertheless, the results of this study supply a warning: higher scores when working together with AI can well “coexist” with worse subsequent unaided-by-AI performance, and the specific type (design) of AI used in the classroom does change educational outcomes.
At work, assistance may become expertise.
A different pattern emerged in a study of 5,172 customer support agents published in the Quarterly Journal of Economics. The researchers examined the rollout of a generative AI assistant that offered customer service representatives suggested responses during live conversations with customers. On average, access to this tool increased the number of issues resolved per hour by 15 percent. Gains were especially strong among newer and less-skilled customer service agents, who had more to learn from the recommendations.
The question is what remained when the AI suggestions were no longer available. The researchers exploited interruptions in the AI assistant’s availability to examine performance without its immediate help. They found evidence that workers exposed to the system retained some productivity gains during outages. The study also suggested that less-experienced workers had begun using communication practices associated with stronger performers.
The findings of this study complicate the notion that every AI-provided shortcut to a better answer erodes competence. The assistant could expose a novice to effective approaches in the middle of real work. Some of those approaches apparently stuck.
As always, we should be precise about the limitations of this study. The researchers studied one type of work in one firm, and the evidence from outages is more suggestive than a randomized long-term examination of retained skill. A customer service script also differs considerably from a correct solution to a mathematical problem or moral reasoning about a complex situation. Still, the result is informative: the same general category of technology that impaired independent performance in one educational setting may have supported durable improvement in another.
A student who copies a complete solution may bypass the very practice the exercise was meant to provide.
What could explain the difference? One plausible hypothesis concerns what the user must do with the advice. A person who revises a suggested response, sees how a customer reacts, and handles the next case may receive useful feedback. A student who copies a complete solution may bypass the very practice the exercise was meant to provide. The studies did not randomize all these mechanisms against each other, so that explanation remains a hypothesis worth testing rather than a definitive conclusion.
A better answer can make judgment worse.
Skill loss is only part of the concern—people also need to recognize when assistance provided by AI is unreliable. An AI tool can improve performance on average, while making some mistakes more likely to survive human review.
In a recent study involving 758 knowledge workers, researchers tested AI on realistic management consulting tasks. For 18 tasks chosen to fall within the AI’s capabilities, participants using AI completed more work, finished faster, and produced higher quality answers. Yet on one complex managerial problem deliberately selected to fall outside those capabilities, the AI-assisted group was 19 percent less likely to reach the correct solution than the group working without it.
This result says something important about judgment at the boundary of a tool’s competence. It does not show that participants had permanently lost professional knowledge. There was no demonstration of deteriorating skill across months of reliance. The experimental finding is narrower and, in a sense, more useful: help that works remarkably well on familiar tasks can steer a person wrong on a deceptively similar one!
Another, slightly older behavioral experiment on trust in AI advice published in 2024, found that participants sometimes followed recommendations they had reason to question, even when the advice conflicted with available information and their own interests. This experiment involved an AI decision aid rather than the more standard AI chatbot. It demonstrated a separate vulnerability produced by overreliance on AI assistance: overreliance can occur at the moment of judgment, without any demonstrated loss of underlying knowledge.
How to Test AI Assistance
Each of these problems discussed here, although on the surface similar, is in fact rather different. A single, average productivity score (“AI improved/decreased performance by X percent”) conceals what’s actually happening. So, how should we measure AI assistance as it becomes more and more commonplace?
Start with assisted performance. Measure accuracy, quality, and time while the AI assistant is available. This establishes whether the AI tool helps deliver an immediate result, which is worthwhile information but only one part of the picture. This is about as far as most (commercial) evaluations go.
Next, measure whether the skill or knowledge is transferable. After AI-supported practice, ask people to complete a related task without using the AI. Compare results with a control group or a credible pre-use baseline, and use delayed tests if retained learning is relevant. For example, a math student should solve a fresh problem rather than recall one shown by an AI chatbot, and a customer support trainee should handle an unfamiliar case without suggested wording for customer interactions. Has unaided judgment improved, held steady, or weakened?
The question for a school or employer is what the person can eventually do, understand, and challenge after repeated AI use.
Finally, test error detection. Offer a plausible AI answer with a known, consequential mistake, and ask people to evaluate it before revealing the defect. Record what they catch, miss, and how confidently they judge it. Establish the correct answer through independent sources or expert reviewers, rather than asking the same AI chatbot to validate itself.
The studies discussed here assessed specific tools and settings, many from earlier AI model generations. Their results are not timeless forecasts. More reliable models will likely reduce errors, but their overall higher quality output might also make occasional mistakes harder to detect. Both possibilities call for continued research and measurement.
Hyoungbin Park is right to worry about the habits cultivated when machines shoulder increasingly complex work. But the currently available evidence does not justify treating all cognitive offloading as damaging. The question for a school or employer is what the person can eventually do, understand, and challenge after repeated AI use. We are yet to know the answers.