Grades dropped from 96 to 48 percent when a Brown professor made students take the exam without AI 布朗大学教授要求学生在无AI情况下参加考试，成绩从96%降至48%

TL;DR

• Brown University professor Roberto Serrano observed a drastic grade drop from 96% on a take-home exam to 48.6% on a proctored final, attributing the initial high scores to widespread AI cheating.
• Two independent studies corroborate this trend: a Chinese study of 26,000 students showed homework scores rising while exam scores fell after AI adoption, and a UC Berkeley analysis of 500,000 grades revealed a significant spike in A grades in unsupervised assignments post-ChatGPT launch.
• The data indicates that reliance on AI tools creates a false sense of mastery, leading to inflated homework performance but a collapse in actual knowledge retention and application during supervised assessments.

Why It Matters

This phenomenon highlights a critical validity crisis in current educational assessment models, particularly those relying heavily on unsupervised, take-home assignments. For educators and institutions, it underscores the urgent need to redesign evaluation methods to distinguish between genuine student capability and AI-generated output, ensuring that grades accurately reflect learning outcomes.

Technical Details

• Brown University Case Study: Professor Serrano identified AI usage by comparing student solutions to ChatGPT outputs, noting that many students utilized complex, non-intuitive mathematical proofs generated by the AI rather than standard direct approaches.
• Central China Longitudinal Study: Tracked 26,000 students (grades 7-12) over 30 months; found that six months after AI introduction, homework completion time dropped from 64 to 45 minutes, homework scores rose by 18%, but exam scores fell by 20%.
• UC Berkeley Grade Analysis: Analyzed over 500,000 grades at a major Texas research university; found that courses with heavy writing/programming components saw a 13 percentage point jump in A grades after ChatGPT’s launch, with the effect being 16 percentage points higher in unsupervised homework compared to proctored exams.
• Performance Disparity: In the Chinese study, approximately 81% of long-term AI users exhibited the pattern of fast/high homework but poor exams, with top-performing students suffering the largest relative declines (24%).

Industry Insight

Educational institutions must pivot toward supervised, in-person, or strictly monitored assessment formats for high-stakes evaluations to maintain academic integrity. Curriculum designers should consider reducing the weight of unsupervised take-home assignments in favor of process-oriented grading or oral defenses that verify individual understanding and prevent AI dependency from masking skill deficits.

TL;DR

• 布朗大学经济学教授发现，无监督家庭作业平均分高达96%，而在监考考试中骤降至48.6%，证实大规模AI作弊现象。
• 两项大型研究佐证了这一趋势：中国一项涉及2.6万名学生的研究显示，使用AI后作业分数上升18%但考试分数下降20%；加州大学伯克利分校的研究显示，ChatGPT发布后，高写作/编程课程的A级比例激增13个百分点。
• 长期依赖AI工具导致学生实际能力显著退化，尤其是优等生表现下滑最严重，且这种负面影响在入学考试中体现为18%-24%的长期损失。

为什么值得看

这篇文章揭示了AI工具在教育领域造成的“虚假繁荣”与“能力空心化”危机，为教育工作者和管理者提供了关于评估体系失效的实证数据。它强调了从依赖无监督作业转向强化监考评估的紧迫性，并警示了过度依赖AI可能导致的社会层面认知能力下降的风险。

技术解析

• 布朗大学案例：教授Roberto Serrano通过对比发现，学生在无监督考试中使用了类似ChatGPT生成的复杂数学证明而非直观解法，监考后平均分从96%暴跌至48.6%，近半数学生退课或不及格。
• 中国纵向研究：追踪2.6万名7-12年级学生30个月，数据显示引入AI后作业完成时间从64分钟缩短至45分钟，作业分数提升18%，但考试分数下降20%；长期用户中81%呈现高分低能特征，优等生受损最重（表现下降24%）。
• UC Berkeley大数据分析：分析德州某大学超50万条成绩记录，发现ChatGPT发布后，侧重写作和编程的课程A级比例上升13个百分点，其中无监督作业密集的课程增幅比监考考试密集课程高出16个百分点。

行业启示

• 教育评估范式重构：高校需重新设计课程评估体系，大幅降低无监督作业权重，增加监考考试、口头答辩或过程性评估的比例，以剥离AI带来的虚假高分。
• 警惕“能力退化”陷阱：政策制定者和教育机构应关注AI对学生长期认知能力的负面影响，特别是其对高阶思维和问题解决能力的侵蚀，需建立相应的干预机制。
• 学术诚信治理升级：面对普遍化的AI作弊，学校不能仅采取个案处理，而应制定统一的AI使用伦理准则和技术检测标准，防止学术诚信体系崩溃。