Overview
- A multidisciplinary team reported that brief access to a GPT-5–based chatbot improved answers on short reasoning tests, yet performance and willingness to keep trying fell once the tool was taken away.
- In the main experiment with about 350 U.S. participants, access to the bot was cut mid-test, and those users then solved fewer fraction problems and quit more often than a control group with no AI.
- The pattern repeated in a larger replication with roughly 670 people and in a final reading-comprehension task with about 200 participants, showing the effect beyond math.
- How people used the chatbot mattered, as participants who asked for hints or clarification coped better after removal than those who prompted for direct answers.
- The paper is not yet peer-reviewed, and the authors warn of a slow "boiling frog" erosion of motivation from heavy AI reliance while calling for designs and policies that build independence in schools and workplaces.