Ten Minutes With a Chatbot Is Enough to Make You Give Up on Fractions, 1,222-Person Study Finds

Berkeley News ran a story on Friday about a study with a finding that most of us have privately run as an experiment on ourselves: give a person an assistant that answers instantly, take it away, and watch them look at the problem like it owes them money.

The study is “AI Assistance Reduces Persistence and Hurts Independent Performance,” with authors from Carnegie Mellon, MIT, Oxford, UCLA and, per Berkeley News, UC Berkeley’s Center for Human-Compatible AI. It covers three randomized controlled trials with 1,222 participants. The AI-assisted sessions were around ten minutes. The damage, apparently, took about the same time. It takes longer to microwave a burrito.

Some precise framing before the jokes, because the authors are careful. Per the arXiv paper, this is about brief AI assistance and later unaided performance in two narrow domains, fractions and reading comprehension. The tasks were short, the time horizon immediate, and per Actuia’s summary only one model, GPT-5, was tested. The effect was smaller in the stricter replication. Berkeley News says the study was presented in October at the Conference on Language Modeling, while Actuia, citing the April preprint, notes it had not been peer-reviewed at that time. The trials were also run on participants recruited online, who were working on fractions for a study and not, say, a cure.

With that said, here are the numbers.

Experiment 1 had 354 recruited participants, 307 analyzed. They solved fraction problems with or without an AI assistant, then took a test without it. On the test, the AI group’s solve rate was 0.57 versus 0.73 for the control group. Their skip rate, the number of times they gave up on a problem, was 0.20 versus 0.11.

Experiment 2, a replication with 667 participants (585 analyzed) and a stricter design, found 0.71 versus 0.77 on solve rate. The skip rate gap, 0.10 versus 0.07, was not statistically significant, per the paper’s table. Sixty-one percent of the AI group said they used the assistant mainly to get direct answers, and that subgroup had the lowest test solve rate (0.65) and the highest skip rate (0.13) of any group. They also declined from their own pretest baseline.

Experiment 3 used reading comprehension with 201 participants. The solve rate was 0.76 versus 0.89, and the skip rate 0.08 versus 0.01.

That last number is the one to sit with: the AI group skipped questions at eight times the rate. Actuia says as much. If you wanted an illustration of “learned helplessness” that included a spreadsheet, here it is.

The authors’ explanation, per the paper and Berkeley News, is conditioning. The assistant gives immediate answers, so slower work starts to feel inefficient, and people lose the experience of working through a challenge, what Berkeley News’s account of the research calls “productive struggle.”

Think about what this describes. Ten minutes in a room with a machine that answers every question quickly, and the human learns a new reflex: if it doesn’t resolve in under four seconds, something is wrong. It is the same lesson we teach to people who have been introduced to a microwave, and the same one that makes them stand in front of it.

The paper’s design proposal is sensible. Brian Christian, the Berkeley researcher featured in the piece, argues AI should be built to be more instructive, like a tutor, instead of defaulting to a quick answer. He is also concerned about overreliance in science, specifically the loss of close engagement with data and deep immersion in material. Berkeley News notes that frontier AI companies have acknowledged that offloading tasks to chatbots can reduce a person’s understanding and mastery of a subject.

So the industry’s own position is: this can happen, and here is a product that does it. This is not hypocrisy exactly. It is closer to the warning label on an energy drink, printed on the side of the energy drink, in a very small font.

There is a fair point on the other side, and the paper concedes it. AI improved performance while available. That is the entire reason people use it. Nobody is claiming the assistant failed to help. The claim is that the help was financed by a loan against future performance, and the interest was charged in skipped questions.

And there’s an obvious practical response, which is that a lot of real work does not involve the AI being taken away. If you have the tool at your desk every day, the “unaided” test is a bit like checking how well a driver navigates without a map after the maps app has been confiscated by a researcher. Those tests measure something, but it isn’t the day job. The authors’ own limitations section (as summarized by Actuia) is open about the narrow tasks and single model.

Still, the persistence result stands apart from accuracy. Getting a fraction wrong is a skill problem. Not trying is something else, and the study suggests that the willingness to try, which is the part of learning that you cannot outsource, is the part that erodes first.

If there is a practical lesson, it is the unglamorous one the researchers point toward: use the assistant like a tutor, not a vending machine. Ask it to explain; try first; keep a minute of friction in the day. This is the AI-era version of “take the stairs,” and it is about as popular.

In the meantime, if you are reading this on your phone and thinking about skipping to the last paragraph, the study would like a word. It would like it slowly. And it would like you to do the fractions.

Sources: