10 minutes with AI weakens the ability to see complex tasks through to the end

At a language model conference this week, an updated version of the study was presented: even ten minutes of relying on AI noticeably weakens the ability to concentrate and to see difficult tasks through to the end. The work has already been peer-reviewed, and among its authors is Brian Christian, a researcher at the Center for Human-Compatible AI at the University of California, Berkeley: on October 7, 2026, he gave a talk about the work at Berkeley. A preliminary draft drew attention back in the spring.

To understand how quickly concentration and persistence fall, the authors conducted a series of randomized controlled experiments and recruited 1,222 participants online. The team included researchers from Carnegie Mellon University, the Massachusetts Institute of Technology, the University of Oxford, and the University of California, Los Angeles.

Three experiments with 1,222 participants

In the first experiment, 354 people solved 15 simple fraction problems. One group kept open a chatbot it could ask for the answer, while the other worked on its own. The group with AI initially answered more accurately, but after 12 tasks the assistant was removed, and accuracy immediately fell. In the second experiment with 667 participants, the result was similar: having lost the AI, people made mistakes or abandoned the tasks altogether, while the group without an assistant saw the work through to the end and in the end coped better.

The third experiment, with 201 participants and a text-comprehension task in the SAT format, produced the same picture: after the AI was switched off, persistence and accuracy declined.

The habit of instant answers

The researchers explain this by the habit of instant answers. When a solution comes within seconds, any task that requires time seems inefficient. In addition, relying on AI removes what is called productive struggle, which helps a person find their own strengths and interests. Christian admits that he notices this in himself as well.

He proposes that AI answers by default be more instructive, like a mentor's, rather than handing out a ready-made solution, and says it is worth moving toward a future where human abilities are, where possible, expanded rather than replaced. This applies especially to science: Christian fears that excessive reliance on AI will undermine what makes a scientist good, that is, closeness to the data, immersion in the material, and teamwork. "This is not a story about fractions and SAT problems, this is happening to knowledge and expertise from top to bottom," he says.

Legal status in Russia: Калифорнийский университет в Беркли — the organization’s activities have been designated undesirable in Russia.

Image:

Translation editor: Салтанов А.