A study reported by UC Berkeley found that participants who used ChatGPT for help on fraction problems or an SAT reading task performed less well when the tool was taken away. The researchers say the results raise a practical question for classrooms and workplaces: does getting an answer faster come at the cost of practising how to reach it?
Watch Desk analysis
What happened
The study involved 1,222 people across three online experiments, according to UC Berkeley’s account. In one experiment, participants with ChatGPT access initially did better on fraction problems. After 12 questions, researchers removed the tool; the Berkeley account says those participants then became less accurate or gave up, while the group without AI was more successful at the end. A second, larger fraction-problem experiment and a third using an SAT reading prompt reportedly showed similar patterns.
The account says the updated, peer-reviewed paper was presented this week at the Conference on Language Modeling. Read UC Berkeley’s account of the study.
Why it matters
The reported result is not simply that people can use AI to get answers. It suggests that, in these tasks, performance after assistance was withdrawn mattered too. That distinction is relevant to anyone deciding whether AI should give learners a direct answer, a hint, or help that makes them work through the problem.
The experiments described here concern particular tasks and an immediate change in performance. They do not establish that using AI generally erodes people’s abilities over time.
Our read
AI that helps someone finish a task is not automatically helping them learn it. For teaching, the interesting design question is whether a tool can offer useful support without quietly removing the practice that builds confidence and persistence. Berkeley’s Brian Christian, a co-author, argues for AI that augments human abilities rather than supplants them. That is a worthwhile aim; the evidence here makes the choice of help, not just access to it, the part to examine.
What to watch
- Whether the paper’s full methods and results clarify how the experiments were run.
- Whether similar effects appear in longer-term studies and other kinds of learning.
- Whether AI tutoring tools can support learners without doing the hard part for them.
Discussion spark: Should educational AI give students direct answers when they ask, or default to hints that make them work through the problem?
Sources and evidence
- Using AI for just 10 minutes erodes your ability to persist at hard things – University of California, Berkeley (9 October 2026, 17:17 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.