AI helped participants with maths and reading tasks. But when the tool was removed, they got fewer answers right and skipped more questions than people who had worked without it.

The results come from randomized experiments with 1,222 people recruited online in the United States. The study was presented at the Conference on Language Modeling on October 9. It measured immediate follow-up performance, not whether any effects lasted.

The gap after AI assistance

In the first fraction experiment, participants used an AI assistant for 12 problems, then answered three questions on their own. They got 57% of those final answers right, compared with 73% in the group that had not used AI.

  • Correct answers: 57% for participants who had used AI, compared with 73% for the control group.
  • Questions skipped: 20% after AI use, compared with 11% in the control group.

A second fraction experiment repeated the pattern while accounting for participants’ starting skills and changes to the on-screen interface. A third experiment tested reading comprehension. The researchers again found lower accuracy and more skipping after AI assistance ended.

What the study does not show

The experiments covered specific school-style tasks in short online sessions. They did not track participants over time or show that everyday AI use causes lasting declines in ability.

The finding is narrower, but still relevant: an immediate boost with AI did not carry over to the next unaided questions in these experiments.

The paper, AI Assistance Reduces Persistence and Hurts Independent Performance, raises a practical question for educators and AI developers: should a tool give users quick answers, or guide them through solving a problem? The study did not test which approach leads to better long-term learning.

For now, the measured difference is about what happened right after the support disappeared. A short-term dip is not proof of permanent skill loss, but it is a reason to test whether AI tools help people learn to work independently.