中文导读 马里兰大学教育学院与信息技术部门在2025年秋季学期开展随机对照试验,评估课程内嵌的虚拟学习助手(Virtual Study Assistant,VSA)。约2,379名本科生、30名教师参与。2026年10月公开的结果表明:获得工具权限的班级期末成绩更低,学习管理系统参与度明显下降,约15%获权限学生实际使用过;对首代大学生的估计损失更大。校方强调“仅提供访问”不足以成为教学干预,并据此向教师提供使用建议、暂缓全校推广。本文据该校教育学院新闻稿、IT部门说明与IZA讨论稿改写。
正文 Many colleges hope that artificial intelligence can give every student a personal tutor . A new University of Maryland study suggests that hope needs careful design . In a Fall 2025 trial , students who received access to a course-integrated virtual study assistant ended the semester with lower grades and far less contact with online course materials than peers without the tool .
本段译文 许多大学希望人工智能能给每位学生配上私人导师。马里兰大学一项新研究显示,这一希望需要审慎设计。在2025年秋季学期的试验中,获得课程内嵌虚拟学习助手权限的学生,期末成绩更低,与在线课程材料的接触也远少于没有该工具的同学。
The research team followed 2,379 undergraduates and 30 instructors across many fields . The chatbot , built on GPT -4o , answered questions mainly from each class’s own ELMS-Canvas pages , files , slides , and video transcripts . Results released in October 2026 describe what the authors call the first large randomized test of such a tool across a whole university . Lead author Jing Liu and colleagues compared sections of the same courses so that teaching content stayed as similar as possible .
本段译文 研究团队跟踪了2,379名本科生和30名教师,覆盖多个学科。该聊天机器人基于 GPT-4o,主要依据各班自己的 ELMS-Canvas 页面、文件、幻灯片和视频文字稿作答。2026年10月公布的结果被作者称为此类工具在整所大学范围内的首次大规模随机试验。第一作者 Jing Liu 与同事比较同一门课的不同班级,以使教学内容尽可能相近。
Among those matched sections , access cut final grades by about 0.37 standard deviations and reduced recorded participation in online activities by about 0.90 standard deviations . Page views and active days on the course site also fell . Losses looked larger for first-generation students . Only about 15 percent of students with access used the assistant at least once , yet classes given access still showed weaker outcomes . Survey clues and campus reports also point to fewer office-hour visits and less time with instructors and peers . Provost Jennifer King Rice noted that offering the tool may have signaled wider acceptance of outside AI agents , even when students never opened the university chatbot .
本段译文 在这些匹配班级中,获得权限使期末成绩约降低0.37个标准差,并使用户在线活动的记录参与度约降低0.90个标准差。课程网站的页面浏览量和活跃天数也下降。对首代大学生,损失看起来更大。获得权限的学生中约15%至少使用过一次助手,但获权限的班级整体仍表现更弱。调查线索与校方报告还指向答疑时间减少、与教师和同伴相处时间变少。教务长 Jennifer King Rice 指出,提供该工具可能发出了更广泛接受校外 AI 工具的信号,即便学生从未打开过学校的聊天机器人。
How students used the assistant also matters . About three-quarters of requests sought information or explanations . Far fewer asked for practice quizzes or feedback on their own attempts . Designers had expected tutoring that creates quizzes and guides learners through productive struggle . Instead , many turned to the bot for quick answers . Few instructors switched from direct-answer mode to a tutor mode that withholds the solution and pushes students to think .
本段译文 学生如何使用助手也很关键。约四分之三的请求是在寻求信息或解释。要求练习测验或对自己尝试给予反馈的要少得多。设计者原本期待能生成测验、引导学生经历“有成效的挣扎”的辅导功能。相反,许多人把机器人当作快速找答案的工具。很少有教师从直接给答案模式切换到不直接给出答案、推动学生思考的辅导模式。
The authors do not claim that AI chatbots harm learning by nature . They argue that access alone is not a teaching plan . Without clear guidance for students and instructors , a helpful-looking tool may replace the effort , office hours , and peer talk that support real understanding . UMD has shared best-practice advice with instructors and is pausing a campus-wide rollout while it studies better ways to embed the assistant . The message is practical : AI can support college study , but only when it supplements course materials and human teaching rather than quietly taking their place .
本段译文 作者并不主张 AI 聊天机器人本质上会损害学习。他们认为,仅提供访问并不是一套教学方案。若没有对学生与教师的明确指导,看似有用的工具可能取代支撑真正理解的努力、答疑时间和同伴交流。马里兰大学已向教师分享最佳实践建议,并在研究更好嵌入方式的同时暂缓全校推广。信息很实际:AI 可以支持大学学习,但前提是补充课程材料与真人教学,而不是悄悄取而代之。
新闻来源