AI News AI资讯 18h ago Updated 2h ago 更新于 2小时前 49

Two-year university study finds banning AI from classrooms leaves students worse off 两年大学研究发现禁止AI进入课堂让学生表现更差

A two-year classroom experiment at Vrije Universiteit Amsterdam found that banning AI entirely produced the worst student outcomes, contradicting the researcher's initial assumptions. Students with structured prompt engineering training outperformed all other groups in 2024, but the advantage nearly disappeared in 2025 as AI familiarity became widespread. Uncritically accepting AI suggestions led to errors persisting in drafts, though students demonstrated ability to self-correct on later exams Thibault Schrepel在阿姆斯特丹自由大学开展两年对比实验,发现"完全禁用AI的学生表现最差",推翻了此前"AI只会带来危害"的预设 接受提示工程培训的第三组在第一年成绩显著领先,但第二年优势几乎消失,原因是学生普遍已具备日常AI使用经验 仅靠AI建议但不加引导的第二组虽然初期存在盲从和误用风险,实际考试表现仍略高于无AI组 教授呼吁高校取消一刀切AI禁令、加强师资AI技能培训,并将论文评价转向实践导向与实证研究

70
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • A two-year classroom experiment at Vrije Universiteit Amsterdam found that banning AI entirely produced the worst student outcomes, contradicting the researcher's initial assumptions.
  • Students with structured prompt engineering training outperformed all other groups in 2024, but the advantage nearly disappeared in 2025 as AI familiarity became widespread.
  • Uncritically accepting AI suggestions led to errors persisting in drafts, though students demonstrated ability to self-correct on later exams without formal instruction.
  • The no-AI group experienced "idea exhaustion" within 10–15 minutes, producing mostly minor wording changes rather than substantive legal improvements.
  • Blanket AI bans in education are counterproductive; universities should invest in AI skills training for both faculty and students rather than restricting tool access.

Why It Matters

This study directly challenges the growing wave of institutional AI bans in higher education, providing empirical evidence that exclusion from AI tools harms learning outcomes more than unguided use does. For AI practitioners and educators, it underscores that the critical variable is not whether students use AI, but how they are trained to use it effectively and critically.

Technical Details

  • Three-group randomized controlled trial across two academic years: Group 1 (no AI, n=66 in 2024 / n=164 in 2025), Group 2 (unguided ChatGPT access with embedded suggestions), Group 3 (hands-on legal prompt engineering and accuracy-checking training).
  • Task: Small teams of 4–5 students revised a provision of the EU AI Act within 20 minutes; grading rubric covered substance, clarity, proportionality, and innovation.
  • Assessment included both in-class performance and a take-home exam requiring revision of a separate AI Act provision, plus a multiple-choice test.
  • Sample was limited to students enrolled in an AI law course, introducing potential self-selection bias toward tech-savvy participants.
  • No mechanism existed to verify actual AI usage during the unsupervised take-home exam, a acknowledged methodological limitation.

Industry Insight

  • Educational institutions should shift from prohibition-based AI policies to investment in structured AI literacy programs, as the research demonstrates that training—not restriction—is the primary driver of positive outcomes.
  • The diminishing returns of formal AI training by year two signal that as tool familiarity becomes commoditized, curricula must evolve continuously rather than treating AI instruction as a one-time intervention.
  • The contrast between boosted homework grades and depressed exam performance in related studies suggests that AI integration strategies must distinguish between formative and summative assessment contexts to avoid superficial learning gains.

TL;DR

  • Thibault Schrepel在阿姆斯特丹自由大学开展两年对比实验,发现"完全禁用AI的学生表现最差",推翻了此前"AI只会带来危害"的预设
  • 接受提示工程培训的第三组在第一年成绩显著领先,但第二年优势几乎消失,原因是学生普遍已具备日常AI使用经验
  • 仅靠AI建议但不加引导的第二组虽然初期存在盲从和误用风险,实际考试表现仍略高于无AI组
  • 教授呼吁高校取消一刀切AI禁令、加强师资AI技能培训,并将论文评价转向实践导向与实证研究

为什么值得看

这项为期两年的课堂随机对照实验为"AI是否应该进入教育"提供了少有的实证证据,直接挑战了当前部分顶尖法学院坚持的"先打基础再用AI"禁令立场。对教育管理者、课程设计者和AI伦理研究者而言,它明确指出了三种典型AI使用模式的学习收益差异,并为学术评价改革提供了可参考的数据支撑。

技术解析

  • 实验设计:随机分三组——禁AI组、直接使用ChatGPT修订欧盟AI法案条款组(无引导)、接受法律提示工程与准确性核查训练组;任务为20分钟内改进条文,评分维度包括实质内容、清晰度、比例原则与创新性,并通过选择题与开卷改写两阶段考核
  • 样本规模:2024年66人,2025年164人,均在修读"AI法"课程的学生中招募,存在一定技术熟练度偏向
  • 结果量化:第一年培训组在综合评分特别是开卷任务上明显领先;第二年三组表现趋同,作者将收敛归因于学生对聊天机器人更熟悉、训练边际收益下降
  • 局限性:期末开卷任务无法核验真实AI使用量;样本局限于AI相关课程;未区分不同学科/任务类型下的异质性效应,也不排除练习效应与年份效应混淆

行业启示

  • 一刀切禁用AI往往导致学习成果平均下降,教育机构应转向"结构化嵌入"策略,例如把提示工程、核查流程与负责任使用纳入必修训练
  • AI素养红利随普及而衰减,课程需要持续迭代:从"工具操作"升级为"批判性整合与责任归属",并在作业—考试链条中重新分配AI使用边界
  • 学术评价体系需要同步改革,纯文献综述类产出可被AI低成本替代,建议提高实践型项目、实证研究与过程性评估权重,避免以传统论文形式掩盖能力虚化

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Education AI 教育AI Legal AI 法律AI LLM 大模型 Research 科学研究