---
id: "lib-2026-10-22-supporting-or-replacing-students-reasoning"
type: "library"
title: "Supporting or Replacing Students' Reasoning? Cognitive Offloading in AI-Assisted Proof Learning"
lang: "EN"
org: "arXiv (CMU, HKU)"
source_type: "Paper"
date: "2026-10-22"
week: "2026-W43"
url: "https://nosov.ai/library/lib-2026-10-22-supporting-or-replacing-students-reasoning/"
original_url: "https://arxiv.org/abs/2610.31481"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Carnegie Mellon field study led with Ken Koedinger that combines random assignment of access with course grades and logged use, and reports its causal limits clearly."
summary_basis: "full-text"
original_minutes: "35"
tags: ["work and ai", "learning", "cognitive offloading", "l&d", "experiments"]
---

# Supporting or Replacing Students' Reasoning? Cognitive Offloading in AI-Assisted Proof Learning

Tier A (Strong evidence): Carnegie Mellon field study led with Ken Koedinger that combines random assignment of access with course grades and logged use, and reports its causal limits clearly.

Summary based on the full text.

## In short

An arXiv preprint (22 October 2026) by Eason Chen, Ken Koedinger and colleagues at Carnegie Mellon, with Jionghao Lin of the University of Hong Kong. In an undergraduate proof-writing course, students were randomly given an AI chatbot and a homework-feedback tool either earlier or later. Early access raised homework scores but not exam scores. Once everyone had access, heavier chatbot use went with lower exam scores, while the feedback tool, which required a proof attempt first, showed no negative association. The authors recommend against open-ended chatbots as tutors.

## Key numbers

- 153: students with grade and usage records (75 early access, 78 later access)
- +2.56 points: homework score advantage of early access, adjusted for prior homework (p = .036)
- B = -2.47 (p = .122): early-access estimate on the next exam taken without AI, not significant
- -3.02 points: next-exam score associated with each additional 100 chatbot messages
- 0.62: exam points lost through chatbot use per one-point drop in baseline self-efficacy

## What it reports

- During the staggered period, 59 of 75 early-access students and 0 of 78 later-access students used the tools.
- Early access: homework +2.56 points after adjustment (95% CI 0.17 to 4.94); Midterm 2, taken without AI, showed no significant difference (raw means 69.95 vs 71.52).
- After universal access, every additional 100 chatbot messages was associated with 3.02 fewer points on Midterm 3.
- Students with lower baseline self-efficacy used both tools more. Mediation: a one-point drop in self-efficacy corresponded to 0.62 fewer exam points through chatbot use.
- Use of the attempt-first homework-feedback tool was not negatively associated with exam scores.
- The chatbot's system prompt told it to avoid final answers, but students could still request them and press for direct help.

## How they got it

308 students enrolled; 155 consented and 153 had complete records. Random assignment at the individual level to access after Midterm 1 or after Midterm 2. Regression models on homework and exam scores, count models for usage, and a path model with 5,000 bootstrap resamples for mediation. Dialogue coding and interviews add context.

## Limits

- One proof-based course at one university; consenting students may differ from others.
- The randomised comparison covers both tools together, so their separate effects are not identified.
- Usage results after universal access are observational; unmeasured differences between students may drive them.
- Cognitive offloading is the authors' interpretation; mental effort was not measured. The final exam was excluded for budget reasons.

## Why it matters for companies

The homework-versus-exam split maps directly to workplace training and onboarding: output produced with AI help is a poor measure of what a person can do without it. Tools that require an attempt before giving feedback look safer for skill building than open chat. The self-efficacy finding suggests the people most likely to lean on a chatbot are those with the most to lose from it, which matters for junior staff and for any learning programme that counts tool usage as success.

Original: https://arxiv.org/abs/2610.31481