Spending just ten minutes relying on an AI tool significantly impairs a person's ability to focus on and persist with difficult tasks, according to a peer-reviewed study presented this week at the Conference on Language Modeling (COLM).
The paper, titled "AI Assistance Reduces Persistence and Hurts Independent Performance," was co-authored by writer and researcher Brian Christian, a research fellow at the UC Berkeley Center for Human-Compatible AI, together with scholars from Carnegie Mellon University, the Massachusetts Institute of Technology, the University of Oxford and UCLA. An earlier draft of the work circulated in the spring, but the updated, peer-reviewed version was presented at the conference this week, and UC Berkeley's research office published a detailed summary on October 9. For more context on this story, see our ongoing AI news.
What the Researchers Found
To measure how quickly people's focus and persistence declined, the team ran a series of randomized controlled trials with 1,222 participants recruited online.
In the first experiment, 354 participants were asked to solve 15 basic fraction problems. One group worked without help, while a second group had ChatGPT open on the side and was free to use it for assistance — or even to ask for answers outright. The group with AI access started out more accurate. Then, after 12 problems, the researchers took the tool away. Almost immediately, accuracy in that group collapsed.
The pattern repeated with a second, larger group of 667 participants. When the AI assistance was withdrawn, the AI users got answers wrong or gave up entirely, while the group that had worked unaided stuck with the task and finished more successfully.
A third experiment with 201 participants swapped the math problems for an SAT reading comprehension prompt. The result echoed the earlier trials: when the tool disappeared, so did persistence and accuracy.
"You almost couldn't ask for a clearer story," Christian said in comments published by UC Berkeley.
The 'Human Cost' of Instant Answers
The researchers argue that the problem is not AI itself but the kind of help it defaults to. When a system produces answers in seconds, users grow accustomed to that speed, and anything slower starts to feel inefficient. Relying on the tool also removes what the team calls "productive struggle" — the effortful engagement that helps people discover their own strengths and interests.
"We've built these systems to be helpful, but are they being helpful to us in a way that's actually helping us?" Christian said. "We showed, ironically, that they are often helping us in ways that are kind of unhelpful."
Christian, whose 2020 book "The Alignment Problem" is widely regarded as one of the leading accessible works on AI safety, said the finding supplies evidence for an intuition many people hold. "It's a striking finding, but in a way it supplies ammunition for a story that a lot of us kind of feel in our gut," he said. "I experience it too."
Why It Matters Beyond Fractions
The study lands amid growing concern about cognitive offloading in education. Frontier AI companies themselves have acknowledged that offloading tasks to chatbots can reduce a user's baseline understanding of a subject, according to Berkeley's report. Christian suggests a possible fix: AI responses could default to being more instructive — behaving like a tutor that guides rather than an answer machine that replaces.
"We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted," he said.
His concern extends to professional research. Christian worries that overreliance on AI tools in science could erode exactly the habits that make a good scientist: staying close to the data, immersing oneself in the material, and working closely with a team. A research group of ten that generates the ideas behind a breakthrough, he noted, might shrink to two or three researchers and a chatbot.
"This is not a story about fractions and SAT problems," Christian said. "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."
A Study That Entered the Mainstream
The early findings drew coverage from WIRED and Psychology Today when the preliminary draft appeared in the spring, and The New York Times published a story about the paper and similar research last week, according to UC Berkeley. The peer-reviewed presentation at COLM — one of the field's main venues for language modeling research — now gives the result a formal stamp.
For educators and employers, the message is uncomfortable but concrete: even short exposure to instant AI answers measurably changes how people behave when the answers are no longer available. And for AI companies, the study adds to mounting pressure to design systems that teach as often as they solve.
Christian, for his part, keeps a notebook on his desk and writes in it every day. The practice, he told Berkeley, keeps him grounded — and keeps his own scholarly edge from dulling.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
