// Contents
The problems were fifth-grade material: add, subtract, multiply and divide ordinary fractions. Participants solved them from home through Prolific, a platform where people earn pocket money taking part in research, and were paid $2.60 for the whole thing. Half of them got a chat window with an AI assistant on the right side of the screen. The other half got nothing. After the twelfth problem the chat disappeared, and both groups faced three more problems on their own.
On those three problems, the group that had the assistant solved 57%, while the group without it solved 73%. The people who lost their helper also abandoned problems almost twice as often: 20% against 11%. About ten minutes of working with AI was enough to produce the effect. The next experiment turned up a curious detail: those who asked the chatbot for ready answers fared worst. Among those who asked for hints, no noticeable losses were found.
The experiment was run by Grace Liu of Carnegie Mellon University, Brian Christian of the Center for Human-Compatible AI (CHAI) at UC Berkeley, known for his books on algorithms and AI, and three more scientists from Oxford, MIT and UCLA. The preprint appeared in April, a revised version was presented in October 2026 at COLM, the Conference on Language Modeling in San Francisco, and on October 9 it was covered by Berkeley News .
Twelve problems with a helper, three without
Psychologists measure persistence in a fairly simple way: how long a person keeps working on a hard problem before throwing in the towel. Here it was counted through skips. Any problem could be skipped, and every skip meant the person had decided not to struggle with it. Both groups saw the correct solution after a mistake, so everyone was learning. One group just learned alone, while the other had a companion who could do the whole thing for them at any moment.
In total, 1,222 people from the United States took part across three experiments, and each one closed a gap in the previous one. Maybe people were thrown off simply because the chat vanished: a window was there a moment ago, and now it is gone. In the second experiment, the no-AI group also got a side panel, which showed solutions to problems the participants had already seen. The gap shrank but held, while the difference in skips this time was not statistically significant, and the authors say so openly.
The third experiment checked whether this was something peculiar to arithmetic. Participants answered reading comprehension questions in the style of the American SAT. The pattern held, and former AI users skipped questions noticeably more often: 8% against 1%.
// Share of problems solved without AI: the group that lost its helper vs the control group
Asking for the answer or for a hint
In the second experiment, participants were asked how they had used the chat, and the results were compared by their answers. Once the chat was switched off, those who had asked for ready answers solved 65% of problems, against 77% in the control group.
// Two ways to use the chat
Give me the answer
This is how 61% of AI participants worked in the second experiment. The machine did the hard part, and the person copied the result. When the chat disappeared, these people solved the fewest problems: 65%.
Show me where to start
This is how 27% of AI participants worked. The person got through the hard part alone, with the chat only nudging. No significant losses were found in this subgroup after the chat was switched off.
The authors add an important caveat here. People chose how to use the chat themselves; it was not assigned at random. Perhaps those who ask for ready answers are the ones who would give up sooner even without any AI. So the breakdown by style shows an association, not a proven cause. The main result is still causal: people randomly assigned to the AI group did worse once it was taken away.
Not a story about fractions
“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,” Christian told Berkeley News. “I experience it too.” In his view, this is not a story about fractions and SAT problems: the same pattern applies to everyone, from grade-school students to the leading experts in the world.
The authors describe the mechanism this way: working with AI trains people to expect an instant answer and takes away the practice of pushing through the hard part on their own. Put simply, it resembles a lift. While it works, nobody needs the stairs. When it breaks down, the fifth floor somehow feels higher than it did a week ago.
Similar things have shown up in other experiments. In 2025 the journal PNAS published a study from a Turkish school where nearly a thousand high school students worked on maths with two versions of GPT-4. Both helped during practice. But on the exam, where AI was no longer available, users of the ordinary chatbot scored 17% lower than classmates who had studied without it. For those who had worked with the tutor version, which gave hints instead of answers, that harm nearly vanished.
In January 2026 Anthropic tested the same idea on programmers . 52 developers, mostly junior, learned an unfamiliar Python library. Those who learned with an AI assistant later scored 50% on a test without it, against 67% for the rest. The gap was widest on debugging questions. And again, those who did better were the ones who asked the machine questions to understand the code rather than handing it all the work.
What ten minutes don’t prove
Headlines like “AI makes us dumber” are not what this experiment shows. It measured one thing: what happens in the first minutes after the help disappears. Whether the effect lasts a day, a month or a year of living with a chatbot is something the study did not test, and the authors name this as a limitation.
The stakes for participants were low, too. For $2.60 and thirteen minutes of time, few people will fight over every fraction, and at an exam or at work motivation is very different. The effect is noticeable but not huge: in the second experiment the difference in solved problems narrowed to six percentage points. And the most encouraging finding, the one about hints, rests on a small subgroup. That is 27% of the AI group in the second experiment, roughly 80 people, and the result still needs to be confirmed.
Still, the picture repeats across three independent studies with very different people: schoolchildren in Turkey, programmers, random adults in the United States. Everywhere, those who handed the hard part of the work entirely to the machine fared worst.
A tutor who won’t let you copy
The fix Christian suggests sounds modest. By default, AI responses could work more like a tutor: lead you towards the idea instead of handing over the finished result. He frames the goal more broadly than any setting: “We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted.”
Such modes already exist. On July 29, 2025, OpenAI launched study mode in ChatGPT : it asks guiding questions and walks you to the solution step by step. Claude has a similar learning mode. But you have to switch it on by hand, and in ChatGPT you can switch it off in the middle of a conversation. The ordinary mode still answers at once and in full.
So the choice between “give me the answer” and “give me a hint” falls to the person, and it comes at the worst possible moment, when the problem is urgent. In this story a hint works like a handrail rather than a lift: you still climb on your own legs, which is why the stairs won’t feel any steeper once the help is gone.
