A peer-reviewed study presented at the Conference on Language Modeling reports that even a short period of artificial-intelligence assistance can leave users less willing or able to persist when they must continue without the tool. The research used randomized controlled trials with 1,222 online participants and tested both mathematics and reading tasks.
In the first experiment, 354 participants worked through 15 basic fraction problems. One group could use ChatGPT, including asking it for answers, while a comparison group completed the work unaided. Participants with AI access were initially more accurate. After the researchers removed that access following the twelfth problem, however, the assisted group experienced a sharp deterioration in accuracy.
A second experiment with 667 participants produced a similar pattern. Once assistance disappeared, former AI users were more likely to answer incorrectly or stop trying, while people who had worked without AI continued longer and achieved stronger results near the end. A third trial involving 201 participants used an SAT reading-comprehension prompt and again found reduced persistence and accuracy after AI was withdrawn.
The findings do not establish that every use of AI harms learning, nor do they show a permanent loss of ability. They capture behavior immediately after assistance in specific controlled tasks. But the consistency across three experiments raises a practical design question: whether tools optimized to deliver answers quickly may discourage the effort through which users develop independent competence.
Brian Christian, a research fellow at the University of California, Berkeley's Center for Human-Compatible AI and a member of the research team, framed that effort as “productive struggle.” The team included scholars affiliated with Carnegie Mellon University, MIT, Oxford and UCLA. Its results suggest that the manner of assistance matters, particularly in education and research where understanding the process is part of the objective.
One response could be to configure AI systems more like tutors than answer engines. Instead of immediately completing a task, a system might ask questions, provide hints or require the user to attempt intermediate steps. That approach could preserve some efficiency while keeping the learner engaged with the material.
The study adds controlled experimental evidence to a wider debate over cognitive offloading. Its most immediate implication is modest but useful: strong performance while a tool is available should not be treated as proof that a user can sustain the same performance independently. Educators and product designers may need to test what happens after assistance ends, not merely measure the gains achieved while it is present.
