Just 10 Minutes of AI Assistance Can Impair Human Problem-Solving and Persistence, Study Finds

Brief AI Assistance Triggers Sharp Decline in Human Problem-Solving

A peer-reviewed study presented at the Conference on Language Modeling reveals that spending just 10 minutes relying on AI assistance significantly impairs human problem-solving abilities and persistence with difficult tasks1

1

2

2

. Brian Christian, research fellow at the UC Berkeley Center for Human-Compatible AI and co-author of the study, led a team of scholars from Carnegie Mellon University, MIT, the University of Oxford, and UCLA in examining how quickly cognitive fitness deteriorates with AI use.

The research team conducted three randomized controlled trials involving 1,222 participants online to test the immediate effects of overreliance on AI tools. The findings challenge the assumption that AI assistance purely augments human capabilities, instead revealing a rapid erosion of independent problem-solving skills.

Randomized Controlled Trials Expose Cognitive Decline

In the first experiment with 354 participants, researchers divided subjects into two groups tasked with solving 15 basic fraction problems. One group worked independently, while the other had access to ChatGPT for assistance or direct answers1

1

. Initially, the AI-assisted group demonstrated higher accuracy. However, after 12 problems, when researchers withdrew the AI tool, performance collapsed. The AI users almost immediately stopped solving questions accurately, while the control group maintained steady progress.

A second, larger trial with 667 participants replicated these results. When AI assistance was withdrawn, users either answered incorrectly or abandoned the task entirely, while participants who worked without AI stuck with the challenge and achieved greater success2

2

. A third experiment involving 201 participants used SAT reading comprehension prompts and confirmed the pattern: AI use erodes persistence and accuracy once the tool is removed.

"You almost couldn't ask for a clearer story," Christian said1

1

.

Productive Struggle and the Cost of Cognitive Offloading

The study reveals that overreliance on AI eliminates what researchers call "productive struggle"—the challenging process that helps people discover their research strengths and passions2

2

. When AI produces answers in seconds, users become conditioned to expect that speed, making anything slower feel inefficient. This cognitive offloading fundamentally alters how we approach difficult problems.

"We've built these systems to be helpful, but are they being helpful to us in a way that's actually helping us?" Christian asked. "We showed, ironically, that they are often helping us in ways that are kind of unhelpful"1

1

.

The findings align with growing anecdotal evidence and surveys documenting how AI assistance undermines baseline understanding and mastery of subjects. Even frontier AI companies have acknowledged this trade-off, though the speed and severity demonstrated in this research marks a significant development.

Implications for Research and Academia

Christian, author of the acclaimed 2020 book "The Alignment Problem," maintains a handwritten notebook despite his deep involvement with AI technology. This practice keeps him grounded against the risk of losing his scholarly edge to excessive AI reliance2

2

.

His concerns extend beyond individual cognitive fitness to the future of research and academia. Overreliance on AI tools in science could jeopardize essential qualities that define good scientists: staying close to data, immersing in material, and collaborating as part of a team. When a research team of 10 people developing breakthrough ideas shrinks to two or three researchers plus a chatbot, something fundamental may be lost1

1

.

"This is not a story about fractions and SAT problems," Christian emphasized. "This is something that is happening to human knowledge and human expertise from top to bottom and from grade-school students all the way to the leading experts in the world"2

2

.

Rethinking AI Tool Design for Human Augmentation

The research suggests a need to fundamentally reconsider how AI systems interact with users. Instead of defaulting to quick answers, AI tool design could prioritize instructive responses that function more like tutors, preserving the productive struggle that builds human expertise1

1

.

"We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted," Christian said2

2

.

The preliminary draft of this study generated significant attention when released earlier in 2026, with coverage from WIRED, Psychology Today, and The New York Times. The updated, peer-reviewed version presented this week provides robust evidence for what many have sensed intuitively: brief exposure to AI assistance can rapidly impair independent problem-solving and persistence, raising questions about how we integrate these tools without sacrificing cognitive fitness and human expertise.