Studies Find AI Assistance Makes Skills Vanish
The Experiment That Watched Skills Vanish
There was an experiment that functioned like a simulation of the future. The doctors in it had been trained to spot polyps during colonoscopies, a specialized skill built through practice and performed through a scope inserted into the colon. Polyps are small growths on the inner wall of the colon, and spotting them is the core of the exam. Then an AI tool arrived that helped them do the spotting. For three months, the machine shared the work, day after day. The doctors still performed the exams; the machine helped them with the spotting. Then the tool was taken away, and the doctors went back to scanning with their own eyes alone. What happened next had never been observed this way before. Their polyp detection rate dropped. The result came from a 2025 study, and its authors suggested that the doctors, in the words relayed by Trent Cash, a behavioral scientist at University of Waterloo in Canada, “kind of forgot what to look for.” [1] In a single measured stretch of time, a professional skill had begun to dissolve, and researchers had watched it happen.
The same pattern surfaced somewhere completely different: a high school classroom. Students were learning a new math concept, and some of them worked through practice problems with an AI tool like ChatGPT at their side. The tool could solve the problems for them, which made the practice feel smooth and productive. Then the researchers took the tool away and tested what the students had truly learned. The students who had used the AI performed worse than students who had never touched the tool at all. [2] The group with the supposed advantage ended up behind the group with no help. Economist Alp Sungu of the Wharton School of the University of Pennsylvania and his colleagues reported the result in the Proceedings of the National Academy of Sciences. The finding is the classroom echo of the hospital one. Math problem-solving is a skill of steps and methods, just as polyp detection is a skill of visual recognition. Both skills faded when a machine did the work.
The pattern repeated a third time, with a skill so basic that its loss is hard to imagine: reading comprehension. People in this study used an AI assistant to solve practice SAT questions, the reading passages and answer choices that students face when applying to college. The AI did the reasoning, and the practice felt productive. When the assistant was removed, the participants had a hard time coming up with correct answers. Machine learning researcher Grace Liu of Carnegie Mellon University in Pittsburgh and her colleagues reported the finding in a preprint posted in April, meaning the study was shared online before formal peer review. [4] The researchers noticed something beyond the wrong answers, something they found more concerning. Participants who lost their AI helper were more likely to skip questions than those who had never used the tool. They simply gave up, the team wrote. “These findings are particularly concerning because persistence is foundational to skill acquisition.” The sentence draws a direct line: skills are built on persistence, and persistence was the first thing the AI had quietly removed.
For Trent Cash, these three experiments reduce to a single principle, one he states as a familiar phrase: use it or lose it. He calls the alternative keeping oneself in the cognitive loop. The loop is the mental work that connects a person to a task — the searching, the reasoning, the deciding, the refusing to give up. When AI completes that work for a person, the loop closes around the machine, and that person’s mind steps out. “If you’re not engaging with the cognitive work, you’re not going to learn the skill,” he says. [1] The doctors who stopped searching lost the search. The students who stopped solving lost the solve. The test-takers who stopped persisting lost the persistence. The mind behaves like a muscle in this respect: it keeps what it exercises and releases what it does not.
Different Houses, One Vanishing Act
The striking thing about these studies is that they came from different houses, pursuing the same question without sharing tools or hallways. Cash makes the general case in the journal Trends in Cognitive Sciences, where he and his colleagues argued on July 9 that learned skills may indeed atrophy if taken over by AI. [1] “The evidence is extremely clear that if we offload a specific skill to AI,” he says, “we’re probably not going to retain that skill particularly well.” The math study came from an economist at the Wharton School of the University of Pennsylvania. The reading study came from a machine learning researcher at Carnegie Mellon University in Pittsburgh. The colonoscopy research came out of clinical practice, from doctors working in hospitals. Each group approached the same question from its own angle, and each angle showed the same slope: down. The convergence is what makes the finding hard to dismiss. One team’s result could be a fluke; several independent sightings are a pattern.
Cash and his colleagues sharpen the picture with a distinction that matters for anyone who uses AI. The question is not whether people use AI but what the tool does to their thinking. AI is most likely to weaken a skill when it replaces the mental work needed to practice that skill. The colonoscopy tool replaced the visual search, and the visual search atrophied. The classroom tool replaced the problem-solving, and the problem-solving drained away. The SAT assistant replaced the reasoning, and the reasoning gave way. In each case, the machine had taken over the exact part of the task that required practice. But the evidence also points the other way. Staying mentally engaged while using AI feedback or examples may preserve skills, and perhaps even enhance them. The difference between erosion and growth sits in the design of the tool and the behavior of the user, not in the mere presence of AI.
Given evidence like this, it would be easy to reach for catastrophic conclusions. Headlines have warned that AI is making people lazy, stupid, and possibly steering civilization toward its downfall. Cash thinks those warnings are probably overblown. He pictures the fear as a kind of digital pocketknife, whittling away at human intelligence one small cut at a time. Real damage, perhaps, but slow and limited. He is hard-pressed to believe the doomsday scenario, in which people completely forget how to think. “Humans are resilient,” he says. “I don’t think that [AI is] going to suddenly turn our brains to mush.” The same resilience that built human skills in the first place, in his view, is what will keep them from vanishing entirely. The erosion is real; the apocalypse is not.
Guardrails That Keep the Mind Working
The natural reaction to these findings would be to avoid AI altogether. The researchers point instead to a middle path. Sungu and his collaborators built an AI tool with learning guardrails to test whether help could be designed safely. The tool worked like ChatGPT but was deliberately constructed not to give away answers. It offered students problem-solving hints and encouragement while they wrestled with a new math concept. The students still had to do the thinking; the machine only pointed the way. The guardrails were the whole point of the design. When the students were tested, this guarded approach helped them learn the concept as well as students who had studied the traditional way, with books. [2] No answers, no atrophy. Help without replacement kept the loop intact.
Sungu frames the difference as one between a tutor and an answer machine. A tutor watches learners struggle and guides them through it. An answer machine removes the struggle entirely, and with it the learning. The students in his study still had to put thought into their studying, and that was exactly the point. “When you’re solving a problem, you need to get your hands dirty,” he says. The messiness of working through a new concept is where real learning happens. The phrase captures the whole argument in a single image: learning is effortful, hands-on, and personal. AI respects that effort when it coaches and erases it when it completes. The guardrail tool was a coach; the earlier tool was an answer machine. The same technology can serve either role; the difference is in the design. Two relationships, two outcomes.
Benjamin Lira Luttges, a behavioral scientist at Wharton, tested whether AI could teach by example, which is a stronger claim than merely doing no harm. He and his colleagues taught people how to edit cover letters, a practical skill with real consequences for anyone applying for jobs. One group practiced with an AI tool. Another group received feedback from human professionals. The AI tool did something that sounds strange at first: it simply rewrote the entire cover letter itself. That sounds like the participants were not doing any thinking at all. But the team found the opposite. Instead of submitting the AI-revised draft right away, people sat with it for a while, tinkering with the language. “They were still working,” Lira Luttges says. Then came the real test. The test measured whether the learning had stuck. Participants edited a poorly written cover letter with no help whatsoever, and human evaluators judged which letters would probably secure a job interview. Letters from both groups were equally likely to land an interview, the researchers reported in a 2026 preprint. AI had taught by example, and the teaching survived the removal of the teacher.
The researchers add a note of caution about which tasks deserve this protection. If there’s a skill you value, don’t give it over to AI, Cash advises. “We know skills come from practice and are maintained through practice.” Lira Luttges agrees and turns the advice into a rule of thumb. People should think about what skills they want to protect, whether that is coming up with the perfect way to phrase a sentence or devising a delectable dish using items in their pantry. “You shouldn’t delegate the things that you enjoy and that give you meaning,” he says. “Outsource the crap, not the craft.” [2] The sentence draws the line that runs through all of these studies. The mind is not a storage room that AI can empty while no one is watching. It is a workshop, and the work itself is what keeps it furnished.