People who use artificial intelligence to solve simple problems may become less capable and more likely to abandon similar tasks when that assistance is removed, according to a new study. The research examined how shifting cognitive work to AI chatbots affected performance and persistence in arithmetic and reading-comprehension tasks, finding that participants who had access to AI performed better while it was available but struggled more when they had to work independently.
The study, presented at the 2026 Conference on Language Modeling in San Francisco earlier this month, investigated a process known as “cognitive offloading.” A growing body of research suggests that cognitive offloading can negatively affect long-term learning and critical-thinking skills.
“What is of concern is what cognitive scientists call ‘desirable difficulty,’ the productive struggle that builds skills over time,” said the study’s lead author, Grace Liu, a doctoral student specializing in machine learning at Carnegie Mellon University, in an email to Live Science. “If AI were to remove that conflict on a daily basis, people might get the right answer in the moment, but they would develop independent capabilities that are less robust.”
Testing AI-assisted problem-solving
Liu and her colleagues recruited 1,222 participants for three randomized controlled experiments. The study was published as a preprint earlier this year and is currently undergoing peer review. The experiments involved two sets of mathematical fraction problems and a set of reading-comprehension questions. The participant groups did not overlap, and each experiment took between 13 and 15 minutes to complete.
In the first experiment, 354 participants were randomly assigned to one of two groups. One group could use an AI chatbot, while the other could not. Both groups received the same 12 fraction problems. Participants with AI access could obtain answers quickly and request hints from the chatbot. The researchers then removed the AI assistant without warning and gave everyone a final set of three fraction problems.
Participants in both groups could skip questions. The researchers said they explicitly told participants that incorrect answers would not be penalized, meaning that skipping represented a deliberate decision not to attempt a problem. Participants also received the correct answer for every question they answered incorrectly, giving them an opportunity to learn from their mistakes.
While the chatbot was available, participants in the AI group were more likely to solve the fraction problems and less likely to give up. Once the AI was removed, however, their performance declined, and they skipped or failed more questions. The findings suggest that using AI impaired unassisted performance, persistence and motivation.
The second experiment largely replicated the first with 667 participants, while making several small changes to the study design. These included a pretest intended to exclude participants who could not solve the most basic questions. The results were similar, suggesting that using AI for simple tasks can reduce people’s perseverance and motivation when they later have to complete the same task without assistance, particularly when the AI is used to obtain a straightforward solution.
Liu acknowledged that calculators also allow people to offload calculations, potentially producing a similar decline in independent performance. “The reason we study AI in particular is because it allows people to offload reasoning across tasks across many domains,” she said.
Similar effect in reading comprehension
In the third experiment, the researchers examined whether AI assistance had a similar effect on reading comprehension. They focused on meaning-making and mental-model building, a cognitive process through which people create internal representations of external reality.
The experiment involved 201 participants divided into an AI-access group and a control group. Both groups first answered five comprehension questions. The researchers then disabled the chatbot and asked each participant three additional questions.
The results were consistent with those of the first two experiments. In particular, they suggested that the reduction in persistence was not an artifact of solving math problems but could be a broader consequence of offloading cognitive work to AI.
“The most important finding is that people who offload more have a greater decrease in persistence (increase in skip rate),” Liu said. “This is an important skill that cuts across domains.”
The importance of productive struggle
One possible reason people stop trying after experiencing AI assistance is that they become accustomed to receiving answers quickly and begin to view slower methods as inefficient, the researchers said. Previous research has shown that perseverance in the face of difficulty is important for overcoming challenges, acquiring skills and learning.
Scientists still do not have a complete picture of how small instances of cognitive offloading affect sustained, independent abilities. Establishing the long-term effects would require researchers to follow the same people over extended periods and regularly measure learning outcomes across a wide range of tasks. However, previous research suggests that overreliance on AI can impair cognitive abilities and mastery of subject matter.
To support long-term learning, competence and autonomy, the researchers argued that AI models and chatbots should guide users toward solutions rather than provide the solutions themselves. In that role, AI would function more like a tutor.
“We need to work towards a future where human capabilities are enhanced as much as possible, rather than replaced,” said computer scientist and researcher Brian Christian of the Center for Human-Compatible AI at the University of California, Berkeley, in a statement.
Liu, G., Christian, B., Dumbalska, T., Bakker, M.A., and Dubey, R. (2026). “AI assistance reduces persistence and impairs independent performance.” Proceedings of the Conference on Language Modeling (COLM). arXiv:2604.04721 [cs.AI].
Source: www.livescience.com