Ten Minutes of Answers: How AI Help Erodes Persistence
A 1,222-person study found minutes of answer-giving AI made people quit hard problems sooner. Hint users held up. The shape of the help is the whole story.
Nobody knows yet whether years of leaning on AI make a person worse at thinking. The long studies haven’t been run. Anyone who gives you a confident answer to that question is guessing.
What we do have now is a much smaller measurement, and it is uncomfortable because it is so small. It comes from a paper first posted as a preprint earlier this year, which Berkeley News reported on today. Its authors are Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel Bakker and Rachit Dubey, and its title is its finding: “AI Assistance Reduces Persistence and Hurts Independent Performance.” Christian is a research fellow at UC Berkeley’s Center for Human-Compatible AI. Across randomized trials with 1,222 online participants, roughly ten to fifteen minutes of AI help on fraction problems and SAT-style reading questions improved how people did while they had it, then left them solving fewer problems, and giving up on more of them, once the help was taken away.
The detail that matters most sits further down. People who used the assistant mostly for hints did about as well as people who never had it.
So the short answer, for anyone who came here with the question: in this study, AI help hurt persistence mainly when it handed over answers. Hint-shaped help looked close to harmless. The rest of this post is about why that split exists and what to do with it.
One evening of dosage practice
Picture a nursing student the night before a medication-math quiz. Dosage problems are fractions wearing scrubs: ordered dose over stock dose, times volume. She has an assistant open in another tab.
Every answer she read was correct, and she followed each one. That is the trap. Following a solution feels like solving it. When the quiz arrives and problem nine doesn’t yield in the first thirty seconds, she has spent the whole evening practicing one response to that feeling: look for the answer. There is no answer to look for, so she moves on. Then she does it again.
She didn’t get worse at fractions in ten minutes. She got practiced at a reflex.
The model: help that takes your next step, and help that leaves it
Here is the model the rest of this post uses, in one sentence. Every piece of help either takes your next step for you or leaves that step with you, and persistence is built only in the second case.
The authors’ own explanation is close to this. They suggest that AI help conditions people to expect immediate answers, and in doing so takes away the experience of working through a difficulty alone. Persistence is the habit of staying in the stuck moment long enough for it to pay off. Answer help deletes the stuck moment. Hint help keeps it and makes it shorter.
The colors in that figure carry through the rest of the post: cobalt for answers, amber for hints and your own work.
Finding 1: ten minutes was enough to move behavior
The design is simple. Participants worked fraction or reading problems, some with an AI assistant and some without. Then the assistant went away and everyone kept going alone.
Accuracy dropping is the expected half. Of course people who had help score lower when it disappears; they practiced less. The skip rate is the strange half. A skip is a decision to stop trying, and the assisted group made that decision more often after only a few minutes of exposure. The group that never had AI held steady.
That is why the paper leads with persistence rather than performance. The assistant didn’t only fail to teach fractions. It changed what people did when a fraction got hard.
Finding 2: the drop followed the answers, not the assistant
The paper then looked inside the AI group, at how people had actually used it. Some asked mostly for direct answers. Some asked mostly for hints.
This is the finding that turns a scary headline into a usable one. If AI in any form made people quit, the advice would be to stay away from it. The data points somewhere narrower. The cost showed up where the tool did the person’s next step, and mostly didn’t where it pointed at that step and stopped.
The authors are careful here, and so should we be. Nobody was assigned to hints. People picked them, and the kind of person who asks for a hint may already be the kind who keeps going. That caveat is real. It is also why the next finding matters, because a different team did run the hint condition as an assigned experiment.
Finding 3: a classroom found the same split, with the hints assigned
Hamsa Bastani and colleagues published “Generative AI without guardrails can harm learning” in PNAS in 2025. They gave nearly a thousand high school math students one of two GPT-4 tutors during practice sessions, or no AI at all, and then tested everyone without it.
The unguarded tutor made practice look great and left students worse off on their own; the working paper put the exam drop at 17 percent. The tutor that was given the solutions but told to give incremental hints instead produced even bigger practice gains, and no measurable exam harm. Same model underneath. Different shape of help.
Put the two studies side by side and the self-selection worry gets smaller. In one, people chose hints and held up. In the other, hints were imposed and students held up. The common factor is that the next step stayed with the learner.
The hint ladder
The practical version of this model is a ladder. You ask for the lowest rung that gets you moving, try again, and only climb when you are actually stuck again.
The ladder is not a rule against answers. Rung four exists. Sometimes the answer is what you need, because the task is something you only need done, or because you have tried and the worked solution is now the fastest teacher. The point is the order. An answer that arrives after your own attempt lands on a question you have already wrestled with. An answer that arrives first replaces the wrestling.
You can set this up in plain language. Tell the assistant you are practicing, that it should start with a question back, and that it should not show a solution until you have written your own attempt. Bastani’s tutor was built on essentially that instruction.
What this gets wrong
The persistence study is careful, and its limits are exactly where a reader is most tempted to over-read it.
The sample was online participants doing short, low-stakes tasks. A paid participant skipping a fraction problem is not a nurse skipping a dosage check, and the motivation in the room is different. The tasks were narrow: arithmetic and reading comprehension, not the messy, open-ended work most adults use AI for. Nobody has shown this effect holds for writing code with an assistant, and coverage of the study has been quick to assume it does.
The window was also short. The effect was measured right after the assistant was removed. That tells us ten minutes of answers can change behavior for the next stretch of problems. It does not tell us whether that change lasts a day, compounds over a semester, or washes out by morning. Both are plausible. Neither is in the data.
And the hint result inside this paper is correlational. Bastani’s classroom study narrows that gap because it assigned the hint tutor, but it is one study, in one subject, with one population. Two studies pointing the same way is a reason to act, not a settled law.
The next step stays yours
Christian’s own suggestion, in the coverage, was that assistants could default to being more instructive, like a tutor, instead of defaulting to quick answers. That is a product decision, and nobody reading this controls it.
What you do control is the first line of the chat. The study’s most useful result is that the cost was not in using AI. It was in letting the AI take the step you were there to practice. Keep that step, ask for the lowest rung, and the same ten minutes that taught the answer-seekers to skip can teach you to stay.
