People who rely on artificial intelligence (AI) to solve simple problems do worse at, and are more likely to give up on, similar tasks when that assistance is removed, compared with people who do not use AI for quick answers, new research finds.
The study, which was presented at the 2026 Conference on Language Modeling in San Francisco earlier this month, investigated the impacts of shifting mental work onto an AI chatbot across arithmetic and reading comprehension tasks. A growing body of research suggests this process of “cognitive offloading” harms long-term learning and critical thinking abilities.
“The concern is about what cognitive scientists call ‘desirable difficulties’ — the productive struggle that builds skill over time,” study first author Grace Liu, a doctoral student in machine learning at Carnegie Mellon University, told Live Science in an email. “If AI routinely removes that struggle, people may get the right answer in the moment but develop less robust independent capability.”
Latest Videos FromLive Science
Liu and her colleagues recruited 1,222 participants for the study, which was released as a preprint earlier this year and has now undergone peer review. The researchers designed three randomized controlled experiments, including two sets of mathematical fraction problems and a series of reading comprehension questions. The participants were split into three groups, so there was no overlap between the experiments. Each experiment took 13 to 15 minutes to complete.
In the first experiment, the team randomly assigned 354 participants to two groups: one that had access to an AI chatbot and one that didn’t. The researchers gave both groups the same 12 fraction problems, which the AI group could get the answer to immediately or obtain clues about by prompting the chatbot. Then, the researchers removed the AI assistant without warning for a final set of three fractions, which all of the participants were given, regardless of their group.
Both groups had the option to skip problems. The researchers explicitly told the participants that there was no penalty for wrong answers, so skipping a problem was a deliberate decision to not engage with it, they wrote in the study. The participants in both groups were also given the correct answer to each problem if they got it wrong, so they could learn from their mistakes.
This first experiment showed that the participants in the AI group were more likely to solve fractions and less likely to give up when they had access to the chatbot compared with the other group. But when the AI was removed, their performance dropped and they skipped or failed more problems, suggesting that using AI impaired unassisted task performance, persistence and motivation.
Get the world’s most fascinating discoveries delivered straight to your inbox.
The second experiment was a copy of the first experiment, but it had 667 participants and small tweaks in the trial design, including the addition of a pretest phase that helped the researchers eliminate participants who were unable to solve the most basic problems. The findings were similar to those of the first experiment, hinting that the use of AI for simple tasks harms perseverance and drive to complete those same tasks without AI, especially when AI is used to get direct solutions.
Liu acknowledged that calculators also allow people to offload computations, which could cause similar declines in independent performance. But “the reason we study AI specifically is that it allows people to offload reasoning across tasks across many domains,” she said.
Calculators allow people to offload math problems, but their effects on reasoning aren’t as far-reaching as the effects of AI.
(Image credit: Kiyoshi Ota/Bloomberg via Getty Images)
In a third experiment, the team explored the impacts of AI assistance on reading comprehension, which draws on meaning-making skills and mental-model construction — the cognitive process of building internal representations of external reality. The trial included 201 participants, who were divided into a group with access to AI and a control group. The researchers gave both groups five comprehension questions before disabling the AI chatbot and handing each participant another three questions.
The results mirrored those of the first two experiments, suggesting that reduced persistence, in particular, was not an artifact of the math problems but a general outcome of cognitive offloading onto AI.
“The most significant finding is that people who offload more have higher declines in persistence (increased skip rate), which is an important skill that cuts across domains,” Liu said.
One possible reason people stop trying once they have experienced help from AI is that they get used to the immediacy of answers and perceive any method that takes longer as inefficient, the researchers wrote in the study. But studies show that persistence through difficulty is essential for skill acquisition and learning, raising concerns about the long-term impacts of using AI for simple tasks.
Scientists don’t yet have a full picture of how small instances of cognitive offloading affect lasting competence and independent ability. To find out, researchers would need to track the same people for a long time and regularly measure their learning outcomes across a wide range of tasks, Liu said. But research so far suggests that an overreliance on AI erodes our cognitive fitness and mastery of subjects, according to the study.
To encourage long-term learning, capability and autonomy in people, AI models and chatbots should be designed to give directions and instructions toward a solution, rather than the solution itself — more like a tutor, the researchers argued.
“We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted,” study co-author Brian Christian, a computer scientist and researcher in the Center for Human-Compatible AI at the University of California, Berkeley, said in a statement.
Liu, G., Christian, B., Dumbalska, T., Bakker, M. A. and Dubey, R. (2026). AI assistance reduces persistence and hurts independent performance. Proceedings of the Conference on Language Modeling (COLM); arXiv preprint arXiv:2604.04721 [cs.AI].
Source: www.livescience.com
