Short AI help made people give up sooner once it was taken away, study finds

The researchers ran three randomized experiments with no overlap between groups: two using fraction problems, with 354 and 667 participants, and one using reading comprehension questions, with 201 participants. In each case, participants with AI access performed better and gave up less while the AI was available. The chatbot was then removed, and everyone, regardless of group, answered a final small set of problems without help.

The article explicitly describes this removal as without warning for the first experiment. The second experiment is described only as a copy of that design, and the third only as disabling the chatbot before the final questions, without the same phrasing repeated. The broad pattern, that previously assisted participants skipped or failed more once unaided, held across all three experiments, based on how the source describes outcomes.

Study first author Grace Liu, a Carnegie Mellon doctoral student, said AI differs from calculators because it lets people offload reasoning across many domains, not just computation. The source frames the reading comprehension result, which mirrored the math findings, as evidence that the persistence drop is a general outcome of offloading onto AI rather than something specific to arithmetic.

Reviewers reading the study draw slightly different conclusions from this point. Some tie the biggest risk specifically to using AI for direct answers rather than hints, pointing to the second experiment's suggestion that perseverance was especially harmed by direct solutions. Others frame the risk more broadly, as following from AI's reach across many kinds of reasoning tasks, without centering how the AI was used. The source itself does not rank which framing is correct, and does not identify specific high-risk occupations or domains beyond the two tested.

The study's authors argue that AI tools should be redesigned to give directions toward a solution rather than the solution itself, acting more like a tutor, to support long-term learning and autonomy. Co-author Brian Christian framed the aim as keeping human abilities augmented rather than supplanted. The source does not report testing whether this redesign actually prevents the drop in persistence seen in the experiments.

Each experiment lasted only 13 to 15 minutes and involved simple fraction or comprehension tasks with no penalty for wrong answers. The source does not describe or test workplace AI tools, longer or more complex tasks, or settings with real performance stakes. Liu said that establishing whether brief AI use causes lasting rather than short-term drops in persistence would require tracking the same people for a long time and regularly measuring learning outcomes across a wide range of tasks. The source states plainly that this full picture does not yet exist.