青少年为完成学业使用 AI 聊天机器人：基于 PISA 2025 的跨国潜在类别分析

英文标题/原题

Adolescents' AI chatbot use for schoolwork: A cross-national latent class analysis of PISA 2025

作者

刘康; 杨晓娟

通讯作者

刘康

研究方向

教育测量；人工智能教育；跨国比较

关键词

generative artificial intelligence; chatbots; PISA 2025; latent class analysis; measurement comparability; negative control; school opportunity to learn; international large-scale assessment

投稿/发表信息

发表期刊/目标

Large-scale Assessments in Education

投稿/成果状态

投稿中

投稿日期

2026-10-22

发表日期

-

原地址/投稿系统地址

https://submission.springernature.com/new-submission/40536/3

期刊与检索情况

检索情况

JIF 3.0（2025），5 年 IF 3.7（来源：期刊主页，2026-10-22 查看）；收录库未核实

摘要/说明

Background: Debates about generative artificial intelligence in schooling often contrast two ways of using chatbots: to understand material or to outsource the work. PISA 2025 asked 15-year-olds how often they used AI chatbots to summarise texts, do preliminary research, draft assignments and get help to learn. We asked whether these four reports form distinct, internationally comparable patterns of use, how the patterns relate to science achievement and to a pre-AI marker of earlier difficulty, and whether schools that teach students to assess AI-generated information change these relations. Methods: In a preregistered analysis, we fitted weighted multigroup latent class models to 627,076 students in 83 education systems. Class enumeration followed fixed admissibility rules, and cross-system comparability was tested with homogeneous, partially homogeneous and heterogeneous models. Achievement associations were estimated within schools, using all plausible values and replicate weights, and pooled with random-effects meta-analysis. Grade repetition in primary school served as a negative control outcome. Results: Six classes were selected. One class, 33% of students, used chatbots mainly to get help to learn. The other five, ranging from non-users to near-daily users of all four purposes, differed in overall intensity rather than in purpose, and no class showed a profile of predominantly outsourcing use. The structure was approximately comparable across systems. Because no outsourcing class emerged, the preregistered class contrast could not be estimated. In prespecified item-level robustness models, a one-step shift toward using chatbots to learn rather than to summarise or draft was associated with 0.117 standard deviations higher science achievement within schools, positive in 76 of 77 systems. The pre-AI repetition marker showed only a small association in the same direction, below the preregistered threshold for a selection signal. School opportunities to assess AI-generated information did not measurably moderate either the socio-economic gradient or the achievement contrast. Conclusions: Across education systems, adolescents' AI use for schoolwork is organised mainly by how much they use chatbots, with a single learning-oriented exception. What students use chatbots for still matters: the balance between learning and producing schoolwork shows a consistent, modest association with achievement that selection on earlier difficulty does not readily explain.

基金信息

无

论文备注

项目目录：E:\论文选题_本地数据集\P08_谁在用AI学习_全球地图。交付：05_稿件/manuscript.pdf、manuscript_blinded.pdf（匿名稿）、cover_letter.pdf、中英对照/。预注册 https://osf.io/rm2gn（2026-10-21）；OSF 项目 https://osf.io/t24dj（公开）。审稿用匿名只读链接：https://osf.io/t24dj/?view_only=781da755fc05425185fb677d91b2143a ；https://osf.io/rm2gn/?view_only=90895c66a0ae421da9391f891dd4aacf