远程教育杂志2026，Vol.44Issue(1)：31-41,50,12.DOI:10.15881/j.cnki.cn33-1304/g4.2026.01.004
大语言模型能否胜任学校教学?
Can Large Language Models Teach in Schools?An Analysis of GPT-4's Coverage Potential and Heterogeneity Across Teacher Competencies
摘要
Abstract
Whether large language models(LLMs)can teach in schools is a pressing practical question in the advancement of educational intelligence.Focusing on LLMs'potential to meet teacher competency requirements,this study first uses GPT-4's perfor-mance on U.S.school assessments across educational levels to identify the knowledge and skills it can plausibly perform in school teaching.These skills are then mapped onto a full-sample teacher job-skill lexicon constructed from U.S.recruitment big data,in or-der to evaluate GPT-4's teaching-capability potential.The results show that skills classified as GPT-4-advantaged achieve an overall coverage rate of 25.2%of the teacher skill lexicon,and the covered skills are disproportionately concentrated in tasks with higher cod-ifiability.In contrast,educational skills that GPT-4 is less able to perform exhibit markedly lower codifiability.Further regression analyses reveal heterogeneous patterns in skill coverage along three dimensions.First,at the subject level,teacher job skills in STEM fields are more likely to be covered by GPT-4,although the coverage gap between STEM and non-STEM teachers is narrowing.Sec-ond,at the educational stage level,higher educational stages are associated with greater coverage;the coverage rates rank from highest to lowest as higher education(including graduate level),high school,middle school,elementary school,and preschool/early childhood.Third,at the task type,broader generalist roles—those requiring a larger number of skills—are more readily covered,whereas more specialized roles—those requiring deeper mastery of specific skills—are less likely to be covered.Overall,the findings suggest that LLMs'aggregate coverage potential delineates a spectrum of teacher skills with codifiability as a shifting boundary.GPT-4's blind spots in expertise-intensive tasks and complex educational practice point to directions for future teacher professional development.Meanwhile,the temporal patterns implied by LLM coverage across educational stages may offer insights for re-examining and adjust-ing the rhythms of schooling.关键词
大语言模型/GPT-4/美国招聘大数据/技能覆盖/人机协同/智能教学/教师专业发展Key words
Large language models/GPT-4/U.S.recruitment big data/Skill coverage/Human-machine collaboration/Intelligent teaching/Teacher professional development分类
社会科学引用本文复制引用
王思宇,陈恺哲,刘进,吕文晶..大语言模型能否胜任学校教学?[J].远程教育杂志,2026,44(1):31-41,50,12.基金项目
本文系国家自然科学基金2023年立项面上项目"'帽子'政策促进还是抑制了学术人才回流?——基于对580万份简历大数据库的人工智能(准)因果推断"(项目编号:72374023)的研究成果. （准）
