Scholarlink

<p>This meta-analysis synthesized empirical evidence on the effects of generative artificial intelligence (GenAI)–supported writing instruction on students’ L2/EFL writing outcomes and examined key moderators. Effect sizes were calculated as Hedges’ g and synthesized using random-effects (DerSimonian–Laird) models. Robustness checks included cluster-robust variance estimation (RVE), sensitivity analyses, prediction intervals, and publication-bias diagnostics (Egger, Begg, PET-PEESE). The primary random-effects model yielded a statistically significant pooled effect (g = 0.80, 95% CI [0.37, 1.22]), indicating a large advantage for AI-enhanced writing instruction. Heterogeneity was substantial (I<sup>2</sup> = 92.21%). Publication-bias tests were non-significant. Robust meta-regression showed that risk-of-bias classification significantly moderated effects (β = 0.93, <i>p</i> = 0.014), whereas instructional model, duration, and sample size were non-significant. AI-enhanced writing instruction demonstrates a positive average effect, though effectiveness appears contingent upon methodological and contextual factors.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Effectiveness of GenAI in enhancing writing performance: a meta-analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Mark Feng Teng

摘要

This meta-analysis synthesized empirical evidence on the effects of generative artificial intelligence (GenAI)–supported writing instruction on students’ L2/EFL writing outcomes and examined key moderators. Effect sizes were calculated as Hedges’ g and synthesized using random-effects (DerSimonian–Laird) models. Robustness checks included cluster-robust variance estimation (RVE), sensitivity analyses, prediction intervals, and publication-bias diagnostics (Egger, Begg, PET-PEESE). The primary random-effects model yielded a statistically significant pooled effect (g = 0.80, 95% CI [0.37, 1.22]), indicating a large advantage for AI-enhanced writing instruction. Heterogeneity was substantial (I2 = 92.21%). Publication-bias tests were non-significant. Robust meta-regression showed that risk-of-bias classification significantly moderated effects (β = 0.93, p = 0.014), whereas instructional model, duration, and sample size were non-significant. AI-enhanced writing instruction demonstrates a positive average effect, though effectiveness appears contingent upon methodological and contextual factors.