Source: ScienceNet
Original title: Study finds: Using AI for 10 minutes reduces human focus
On the desk of Brian Christian, a researcher at the Center for Human-Compatible AI at the University of California, Berkeley, there is always a notebook. This scholar, who has spent two decades contemplating AI, insists on handwriting every day. Underlying this habit is a persistent question: Could working too much with AI dull his academic edge?
His latest research demonstrates that this concern is not unfounded.
At the recent Conference on Language Modeling (COLM), AI researchers announced a peer-reviewed study showing that using AI tools for just 10 minutes may significantly impair people’s ability to focus and persist on challenging tasks.
The study titled "AI Assistance Reduces Persistence and Impairs Independent Performance" was conducted by scholars from UC Berkeley, Carnegie Mellon University, MIT, Oxford University, UCLA, and other institutions. Researchers recruited 1,222 participants online for a randomized controlled trial, whose results reveal an unavoidable question: How should people interact with AI?
When the AI withdraws, performance drops.
Brian Christian has long focused on the impact of AI on the human mind; his 2020 book, The Alignment Problem, is regarded as one of the best in the field. Like many scientists, he is sometimes exhilarated by AI’s progress, but he has always harbored underlying concerns about “erosion.” What he most wants to understand is: how fast and how deep is this erosion?
To this end, the research team designed a series of randomized controlled experiments. In the first experiment, 354 participants were divided into two groups: one group solved 15 basic math problems independently, while the other group had access to the chatbot ChatGPT, allowing them to seek help at any time—even to ask for direct answers. Initially, the group with AI assistance answered more accurately. However, when the AI tool was removed at question 12, the accuracy of this group dropped sharply, and the rate of skipped questions rose significantly.
The second experiment was larger, involving 667 participants, and the results were identical: without AI assistance, former AI users either answered incorrectly or gave up entirely, while the group that had never used AI persisted through the task and ultimately performed better.
In the third experiment, using 201 participants and reading comprehension questions from the U.S. SAT, the results remained the same: when the AI tool was removed, both persistence and accuracy declined.
AI will diminish the enthusiasm for trial and error and perseverance.
These findings corroborate numerous recent investigations and everyday observations: overreliance on AI-generated answers undermines people’s cognitive adaptability. The reason is straightforward: when AI provides answers in seconds, people grow accustomed to this speed and perceive anything requiring more time as “inefficient.” More importantly, AI deprives individuals of what researchers call “productive struggle”—the process of grappling with challenges, experimenting, and persisting, which guides people toward discovering their strengths and passions.
The cognitive science community is no stranger to this. In 2011, psychologists Sparrow et al. published the famous "Google effect" experiment in Science: people who knew information could be easily searched online showed poorer memory for the information itself, remembering only one thing—where to look for the answer next time.
The theory of "desirable difficulties" proposed by UCLA psychologist Bjork also suggests that moderate challenges in learning are not detrimental but rather enhance understanding and long-term memory. In 2014, researchers from Princeton University and UCLA found that students who typed notes on a keyboard had poorer comprehension of the material than those who wrote by hand and were forced to summarize and distill the content.
The convenience brought by tools often comes at the cost of skill acquisition, a finding recently supported by the latest research report. Christian says this is a striking discovery, but in some ways, it simply brings to light what many have long felt deep down.
Earlier this year, the initial draft of this study generated significant attention, with media outlets such as Wired, Psychology Today, and The New York Times reporting on it. The version recently submitted to the conference is the updated, peer-reviewed edition.
AI can be a mentor, but not a substitute.
Interestingly, leading AI companies themselves acknowledge that delegating tasks entirely to chatbots can reduce users' foundational understanding and mastery of a subject.
To reverse this, AI researchers believe the solution lies in redefining AI’s “role”—its responses should gently guide and inspire thinking, like a mentor, rather than defaulting to simply providing answers.
This warning is especially important for the scientific and academic communities. Researchers fear that overreliance on AI tools threatens the very foundation of excellent science: close engagement with data, deep immersion in materials, and collaborative teamwork. They envision a scenario where a breakthrough idea originally pursued by a team of ten might be reduced to just two or three researchers plus a chatbot, potentially losing many truly essential elements.
And this isn't a story about solving math and physics problems faster, Christian says: “This is what’s happening across the board—from elementary school classrooms education to the world’s top experts—in human knowledge and expertise.”
Setting regulations for AI in education has become a global initiative. The European Union’s Artificial Intelligence Act, effective in August 2024, classifies education-related AI applications—such as admissions, academic assessment, and exam monitoring—as high-risk, subjecting them to strict controls. As of mid-2025, more than half of U.S. states have issued guidelines for AI use in public schools.
In China, the education sector is also taking action. In May 2025, the Basic Education Teaching Guidance Committee of the Ministry of Education issued the "Guidelines for the Use of Generative AI by Primary and Secondary School Students (2025 Edition)," clearly requiring that students not directly copy AI-generated content as homework or exam answers, thereby addressing at the source the misuse of AI as a substitute for independent work.
“We should strive for a future that enhances, rather than replaces, human capabilities,” AI researchers emphasize. This research does not seek to deny the value of AI, but rather to remind people that it is the way we use it that determines its benefits and drawbacks.
Twitter: https://twitter.com/BitpushNewsCN
BitPush Telegram community: https://t.me/BitPushCommunity
BitPush TG subscription: https://t.me/bitpush