"AI brain rot" and cognitive offloading AI tutors and learning outcomes Motivation
AI help raised homework scores, but not exam scores, and heavy chatbot use went with lower marks
In a university proof-writing maths course, students were randomly given access to AI help (a chatbot plus a tool that critiques a submitted proof) either earlier or later. Those with earlier access scored higher on homework, but not on the next exam, which everyone sat without AI. Later, students who sent the chatbot more messages tended to score lower on the following exam, and less confident students used it more.
| Paper | Supporting or Replacing Students' Reasoning? Cognitive Offloading in AI-Assisted Proof Learning |
| Authors | Eason Chen, Sophia Judicke, Kayla Beigh, Xinyi Tang, Isabel Wang, Nina Yuan, Zimo Xiao, Chuangji Li and 11 others, with Jionghao Lin and Ken Koedinger |
| Published | 22 October 2026 · arXiv preprint 2610.31481 (cs.HC) |
| Study type | Randomized staggered-access field study plus correlational usage analysis · Preprint (not yet peer-reviewed) |
| What we read | Full text read. We read the full preprint on arXiv (version 1, 22 October 2026). |
| Summarized |
What they did
After the first midterm, consenting students were randomly assigned as individuals to get the two AI tools straight away or only after the second midterm. The researchers compared homework and the next exam between the groups. Once everyone had access, they counted each student's logged chatbot messages and feedback requests, and related them to the third midterm while controlling for prior exam scores and self-rated confidence (self-efficacy). They also coded what students asked for and interviewed 19 volunteers.
Who took part
153 students (75 early access, 78 later access) in an undergraduate course on logic, sets, relations and proof writing at Carnegie Mellon University. The chatbot used GPT-4o with a prompt telling it to give Socratic questions and partial hints rather than final answers.
What they found
- While only the early group had access, their homework scored 2.56 points higher after adjusting for earlier homework (95% CI 0.17 to 4.94).
- On the next exam, taken without AI, the groups did not differ significantly (raw means 69.95 for early access vs 71.52 for later access).
- After everyone had access, each extra 100 chatbot messages went with a 3.02-point lower score on the next exam, after accounting for prior exam score, confidence and use of the feedback tool.
- Students whose confidence score was one point higher sent about 39% fewer chatbot messages and made about 31% fewer feedback requests.
- Using the tool that required a proof attempt before giving feedback was not negatively linked to exam scores (the association was positive but not statistically reliable).
- Chat logs show students pushing the "hints only" chatbot with requests like "Tell me the answer", and it giving increasingly explicit steps. In both example chats the authors highlight, the chatbot also made a maths error.
Effect size: Homework +2.56 points (randomized); next exam −2.47 points, not significant (95% CI −5.60 to 0.66); −3.02 exam points per 100 chatbot messages (correlational).
Limitations
- A preprint, not yet peer-reviewed.
- University students in one proof-based maths course, not children or school homework.
- Only the timing of access was randomized. How much students used each tool was their choice, so the link between heavy chatbot use and lower scores is correlational: struggling students may simply seek more help.
- The two tools were given together, so their separate effects could not be identified.
- Cognitive offloading is the authors' interpretation; mental effort was not measured directly.
What this means for learners and parents
- Better homework with AI is not evidence of learning. Check what a child can do on a similar problem without help.
- A "hints only" instruction is not enough on its own: learners can push a chatbot for more and more of the solution.
- Help that starts from the learner's own attempt (try first, then get feedback) looked better than open-ended chat in this course.
- Children who feel least confident may lean on AI the most, so they need the most attention to whether they are still doing the thinking.
Put it into practice (free tools)
- Is this AI doing your homework? (prompt checker)
- Should my 8 or 9-year-old use ChatGPT for homework?
- Hint ladder: hints instead of answers
Citation
Eason Chen, Sophia Judicke, Kayla Beigh, Xinyi Tang, Isabel Wang, Nina Yuan, Zimo Xiao, Chuangji Li and 11 others, with Jionghao Lin and Ken Koedinger (2026). Supporting or Replacing Students' Reasoning? Cognitive Offloading in AI-Assisted Proof Learning. arXiv preprint 2610.31481 (cs.HC). https://doi.org/10.48550/arXiv.2610.31481
Related summaries
- Unrestricted ChatGPT help raised practice scores but lowered exam scores · Hamsa Bastani et al., 2025 · Preregistered randomized controlled trial
- A carefully designed AI tutor beat an active-learning physics class on a short-term test · Greg Kestin et al., 2025 · Randomized controlled trial with a crossover design
- Supervised AI tutoring matched human tutors in UK maths classrooms · LearnLM Team (Google) et al., 2025 · Exploratory randomized controlled trial
- An AI coach for human tutors raised topic mastery, most for weaker tutors · Rose Wang et al., 2024 · Preregistered randomized controlled trial