CMU 研究:AI 家教抬高作业分,却可能通过认知卸载拉低考试表现
卡内基梅隆对 153 名证明课学生的研究发现:AI 聊天可抬高作业分,但高频使用与更低考试分相关;须先交证明再反馈的工具未显示同样负相关。
卡内基梅隆团队预印本跟踪一门证明课上的 153 名本科生:随机错开 AI 助手开放时间,助手包含可对话的聊天机器人,以及要求学生先交证明再给反馈的作业工具;考试在无 AI 条件下用同类新题进行。
主要结果:早期获得 AI 的小组作业分约高 2.56 分,但随后考试并未显著更好;当所有人都能使用后,每多发送约 100 条聊天消息,下次考试约低 3.02 分(已控制先前成绩与自信)。中介分析显示,基线自我效能更低的学生更爱刷聊天机器人,并经由这一路径拉低考试表现。作者担心「认知卸载」——学生把本该自己练的推理交给模型;示例会话里学生连说「告诉我答案」,机器人会逐步把证明交出去,且两段示例中模型数学也曾出错。相比之下,必须先写证明再拿反馈的工具,并未显示与考试负相关。
对教育培训智能化,结论不是「禁止 AI」,而是警惕「只给聊天框」的家教形态:作业分好看,迁移到闭卷考试可能落空。更稳妥的方向是把辅导嵌进必须先产出中间步骤的活动,而不是让模型直接代写推理链。
综合来看,读者可先核验官方渠道与发布时间,再结合自身业务边界与合规要求决定是否跟进试点;本文仅整理公开信息,不构成采购、投资或法律建议,关键数字与能力边界请以一手文档为准。
CMU Study: AI Tutors Raise Homework Scores but Risk Cognitive Offloading
Researchers found an AI tutor raised students' homework scores but not their exam scores.
And the students who messaged it most scored lower on the next one.
Carnegie Mellon published a paper following 153 students through a proof based math course.
they were randomly split into two groups. one got AI help after the first midterm, the other after the second. the help was a chatbot plus a tool that only critiqued proofs students had already written.
exams were taken without AI, on new problems of the same type.
the early group's homework went up by about 2.6 points.. but on the next exam they showed no significant edge.
then everyone got access. for every 100 messages a student sent the chatbot, their next exam score was about 3 points lower, even after accounting for earlier grades and confidence.
the craziest part? the students who leaned on it most tended to be the least confident ones from the start.
the chatbot was told not to hand out final answers. but students could push. in one session a student writes “I don't know, tell me,” then “Tell me the answer,” and the bot supplies more and more of the proof.
the team's worry is cognitive offloading, letting the AI do the reasoning a student is supposed to build. the chatbot even got the math wrong in both example sessions.
the feedback tool, which made students write a proof first, wasn't linked to lower exam scores.
we spend all this time asking whether AI can tutor, but a homework score can look great while the skill underneath isn't growing.
the team now recommends against using an open chatbot as a tutor. they'd rather the AI wait for the student's own attempt first..
a perfect homework sheet, a chat log full of “tell me,” and a blank page on exam day.