Brief AI Access Lowers Unaided Test Scores

A set of randomized U.S. experiments found that participants given short GPT-5 access performed worse on questions answered without it.

Updated on Nov. 9, 2026 in Homework

Brief AI Access Lowers Unaided Test Scores

Live Poll

Do you plan to use AI tools for work or study tasks this year?

Researchers recruited 1,222 participants across three randomized experiments and gave one group access to GPT-5 during sessions lasting about 10 to 15 minutes. When that access was removed for the final questions, the AI-access group answered fewer questions correctly than control groups in each experiment.

Why it matters

The findings suggest that brief reliance on an AI tool may weaken performance once the tool is no longer available, a concern for students and workers who use such tools on everyday tasks. Because assignment to access did not guarantee use, the results describe group-level effects rather than the outcomes of confirmed users.

In the first experiment, the AI-access group skipped 20% of final questions versus 11% for controls, and answered 57% correctly versus 74%. Sessions lasted about 10 to 15 minutes, and the study's long-term effects remain unmeasured in the source material.

The details

Participants were given GPT-5 beside the questions in sessions lasting about 10 to 15 minutes, during which it could provide complete solutions. Researchers removed access before the final three questions in each experiment and measured performance without it.

Sessions with GPT-5 access lasted about 10 to 15 minutes each.
Final questions were measured immediately after GPT-5 access ended.

Big Picture

The study applies a randomized controlled design to test whether brief generative AI access affects later unaided performance, a question that matters as such tools enter homework and work tasks. Its results are group-level comparisons rather than measurements of confirmed users.

Students who use AI tools for assignments may find that practice on unaided questions matters for later exams. The results suggest checking that core skills stay sharp even when a tool is available for help.

The takeaway

Relying on an AI tool for quick answers may leave unaided performance weaker once it is removed. Practicing key problems without assistance can help keep those skills intact.