Ten minutes of AI help made people give up sooner on hard problems, study finds

A team of researchers led by Brian Christian of the University of California, Berkeley's Center for Human-Compatible AI has found that a short spell of help from an AI chatbot can make people less willing to keep going on hard problems once the help is taken away. The work was presented at the Conference on Language Modeling on October 9, 2026, and publicised by the university the same day. In three randomised online experiments with 1,222 participants in total, people who had an AI assistant for roughly ten minutes solved fewer problems on their own afterwards and skipped more questions than people who never had the assistant. The researchers also found an important exception: people who used the AI for hints rather than ready-made answers were not harmed in the same way.

Most debate about AI in classrooms and offices has been about cheating or accuracy. This study looks at something harder to see: what happens to a person's own effort after the tool has done some of the work. The answer from these experiments is that the cost can show up very quickly, after minutes rather than months, and it shows up as giving up sooner, not only as getting answers wrong.

The finding on hints is the most useful part for readers. It suggests the problem is not AI itself but how it is used. When people asked for the answer, they learned less about how to get there and had less stamina when they were on their own. When they asked for a nudge, they did about as well as people who had worked alone throughout. That matches long-standing teaching practice: a good tutor gives a hint and lets the student finish the step.

The results should be read with care. The tasks were short and done online, the way people used the tool was partly self-reported, and the authors note that the split between answer-seekers and hint-seekers is not proof of cause and effect. The assistant in the trials also gave full answers when asked, so the study does not test tutoring tools built to hold answers back.

For India, where millions of school and college students are preparing for board exams and entrance tests and AI chatbots are a few taps away, the lesson is practical. A student who uses AI to check reasoning or get unstuck may be fine; one who uses it to skip the thinking may find the exam hall harder, because the exam is taken without the tool.

Short-term: Teachers, parents and students get evidence to support "hint first" rules for homework and practice, and companies building AI tutors get a design target.

Long-term: If the effect holds in longer, real-world studies, schools and employers may need to build AI tools and policies that protect independent effort, not just accuracy.

Who is affected: Students and exam aspirants, teachers, edtech companies, employers who train staff with AI tools, and AI developers designing learning modes.

Questions and Answers