A new study from researchers at the University of California, Los Angeles, Massachusetts Institute of Technology, University of Oxford, and Carnegie Mellon University suggests that even short-term use of artificial intelligence tools may weaken people’s ability to solve problems independently.
The research involved 1,222 participants across three randomized controlled experiments designed to test how AI assistance affects learning, persistence, and cognitive performance. Participants were assigned math and reading comprehension tasks, with some given access to a chatbot powered by GPT-5 while others completed the work without AI assistance.
Initially, the AI-assisted participants performed better than the control group. They answered more questions correctly and completed assignments more quickly. But researchers introduced a second phase after only 10 to 15 minutes, removing access to the chatbot and asking all participants to continue on their own.
That is where the results shifted dramatically.
Participants who had previously relied on AI performed significantly worse once the chatbot was removed. In math-related tasks, the AI-assisted group solved only 57 percent of problems correctly, compared to 73 percent among participants who never used AI. Reading comprehension scores also declined, with the AI group scoring 76 percent compared to 89 percent in the control group.
Researchers also found that participants who used AI were more likely to skip difficult questions and abandon challenging problems altogether.
According to the researchers, the most concerning finding was not simply the drop in performance, but the decline in persistence.
“It revealed that AI helped in the moment, but when participants returned to working on their own, they performed worse and were more likely to give up,” the study analysis noted.
The research raises broader concerns about how AI tools could reshape learning habits and critical thinking over time. While AI systems can provide instant answers and save time, researchers warned that repeated dependence on them may reduce the mental effort required to work through difficult problems independently.
Skills such as reasoning, perseverance, and confidence gained through overcoming challenges could gradually weaken if users consistently rely on AI-generated solutions.
Researchers compared the process to the classic “boiling frog” metaphor — the idea that small, gradual changes often go unnoticed until the cumulative effects become significant.
“Each instance of relying on AI feels harmless,” the researchers explained, “but over time the accumulated impact on cognitive effort may become difficult to reverse.”
The study warned that modern AI systems are currently optimized primarily for short-term helpfulness rather than long-term human development.
“We caution that if such effects accumulate with sustained AI use, current AI systems — optimized only for short-term helpfulness — risk eroding the very human capabilities they are meant to support,” the researchers wrote.
The researchers added that the danger extends beyond lower task performance.
“People do not merely become worse at tasks, but they also stop trying,” the study stated. “If such effects accumulate over months and years of AI use, we may end up creating a generation of learners who have lost the disposition to struggle productively without technological support.”
Rachit Dubey, a co-author of the study, warned that rapid AI adoption in education could eventually produce “a generation of learners who will not know what they’re capable of,” potentially reducing creativity and innovation on a large scale.
Rather than abandoning AI entirely, researchers are encouraging developers to rethink how these systems interact with users. Instead of delivering immediate complete answers, they suggest AI tools should act more like tutors or collaborators that guide users through challenges while still requiring independent thought.
“This requires rethinking how AI systems are built to collaborate with humans,” the researchers wrote, “and just as the best human collaborators know when not to help, so too should AI.”
