Study Finds Lower Unaided Scores After Brief Tasks With GPT-5 Access

Access to GPT-5 helped participants during short math and reading tasks, but groups offered the tool scored lower on unaided questions immediately afterward in three randomized experiments involving U.S.-based participants.

The experiments recruited 1,222 people: 354 for the first fraction test, 667 for the second and 201 for the reading test. After exclusions, the analyses included 305, 585 and 168 participants, respectively.

During sessions lasting about 10 to 15 minutes, GPT-5 was available beside the questions and could provide complete solutions. Scores improved while the assistant was available. Researchers removed access before the final three questions in each experiment, then measured how participants performed without it.

In the first fraction experiment, participants assigned AI access answered an average of 57% of the final questions correctly, compared with 74% for those without access. The AI-access group skipped 20% of those questions, versus 11% in the control group.

The second fraction experiment found a smaller accuracy gap. Participants previously offered GPT-5 answered 71% of the final questions correctly, compared with 77% for the control group. Skip rates were 10% and 7%, respectively, but that difference was not statistically significant in the full comparison.

On the final reading questions, the AI-access group answered 76% correctly, compared with 89% for participants who had worked without AI. The groups skipped 8% and 1% of questions, respectively.

Assignment to an AI-access group did not mean every participant used GPT-5. Some people offered the tool reported that they did not use it. The results therefore compare groups given access with control groups, rather than confirmed users with nonusers.

The experiments measured performance immediately after access ended. They did not establish whether the differences lasted for hours or days, whether they would appear in other kinds of work, or whether participants’ intelligence changed.

Grace Liu and her co-authors wrote, “We urge caution in interpreting these results as broad evidence that AI lowers persistence and independent ability for all tasks.”