How AI Assistance Affects Human Skill Development: A Study of Learning with Logic Puzzles
Source: https://arxiv.org/abs/2608.23543
Authors: Shang Wu, Catarina G Belem, Shuyuan Fu, Mark Steyvers, Padhraic Smyth
Published: 2026-08-24 (arXiv 2608.23543v1) · Accepted at Human-AI Complementarity and Alignment (HCOMP) 2026
Category: cs.AI
English Summary
While AI assistance can improve human task performance in the short term, it may also undermine the development of skills in the longer term. The authors examine this tension in a controlled logic-puzzle experiment involving on-demand AI assistance, where participants complete tasks before, during, and after AI is available.
Key findings from the abstract:
- By experimentally varying AI request costs, lower-cost assistance induces more frequent AI use.
- Participants who request AI assistance during the AI-access phase perform worse at the task after assistance is removed, and their subsequent unassisted performance is overestimated when predicted from earlier AI-assisted performance.
- A Bayesian latent ability model separates initial ability, post-AI ability, and participant-specific skill change, while estimating how independent reasoning during the AI-access phase relates to skill development.
- Greater independent problem-solving effort is associated with larger gains in latent ability — consistent with the interpretation that skill development is weaker when AI assistance substitutes for independent reasoning.
中文摘要（要点）
AI \u8f85\u52a9\u53ef\u4ee5\u63d0\u5347\u77ed\u671f\u4efb\u52a1\u8868\u73b0\uff0c\u4f46\u53ef\u80fd\u524a\u5f31\u957f\u671f\u6280\u80fd\u517b\u6210\u3002\u8fd9\u9879 HCOMP 2026 \u7814\u7a76\u7528\u53d7\u63a7\u903b\u8f91\u8c1c\u9898\u5b9e\u9a8c\u76f4\u63a5\u6d4b\u8bd5\u8fd9\u4e00\u5f20\u529b\uff1a\u88ab\u8bd5\u5728 AI \u53ef\u7528\u524d\u3001\u4e2d\u3001\u540e\u4e09\u4e2a\u9636\u6bb5\u5b8c\u6210\u4efb\u52a1\u3002
- \u5b9e\u9a8c\u64cd\u7eb5 AI \u8bf7\u6c42\u6210\u672c\uff1a\u6210\u672c\u8d8a\u4f4e\uff0cAI \u4f7f\u7528\u8d8a\u9891\u7e41\u3002
- \u5728 AI \u53ef\u7528\u9636\u6bb5\u8bf7\u6c42\u8fc7\u8f85\u52a9\u7684\u88ab\u8bd5\uff0c\u5728\u8f85\u52a9\u6491\u9664\u540e\u4efb\u52a1\u8868\u73b0\u66f4\u5dee\uff0c\u4e14\u7528\u65e9\u671f AI \u8f85\u52a9\u4e0b\u7684\u8868\u73b0\u9884\u6d4b\u540e\u7eed\u65e0\u8f85\u52a9\u8868\u73b0\u4f1a\u7cfb\u7edf\u6027\u9ad8\u4f30\u3002
- \u4f5c\u8005\u7528\u8d1d\u53f6\u65af\u6f5c\u5728\u80fd\u529b\u6a21\u578b\u5206\u79bb\u521d\u59cb\u80fd\u529b\u3001\u540e AI \u80fd\u529b\u4e0e\u4e2a\u4f53\u6280\u80fd\u53d8\u5316\uff0c\u5e76\u4f30\u8ba1 AI \u9636\u6bb5\u7684\u72ec\u7acb\u63a8\u7406\u4e0e\u6280\u80fd\u53d1\u5c55\u7684\u5173\u7cfb\u3002
- \u7ed3\u679c\uff1a\u72ec\u7acb\u89e3\u9898\u52aa\u529b\u8d8a\u591a\uff0c\u6f5c\u5728\u80fd\u529b\u63d0\u5347\u8d8a\u5927\uff1b\u5f53 AI \u8f85\u52a9\u66ff\u4ee3\u4e86\u72ec\u7acb\u63a8\u7406\u65f6\uff0c\u6280\u80fd\u53d1\u5c55\u66f4\u5f31\u3002
Why It Matters
This is direct experimental evidence on the "cognitive debt" question for AI-assisted learning: access to capable assistance changes not just output quality but the learning curve behind it. The request-cost manipulation is the most policy-relevant part — making help frictionless increases usage, and heavier usage during the access phase predicts worse unassisted retention. For anyone designing AI tutoring or copilot products, the measurable gap between AI-assisted and later unassisted performance is the metric to watch, and the Bayesian decomposition offers a reusable analysis pattern.