Instructional model shows a significant subgroup difference (traditional classroom g = 1.19 vs. AI teaching g = 0.69) that does not remain robust in multivariate meta-regression¶

Claim · All claims
Evidence · 1 study (2 entries) · quant-synthesis r2–r3 · q3
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q3 i? Subgroup analysis found significant between-group heterogeneity by instructional model (QB = 27.26, p < .001), with traditional classroom implementations (g = 1.19) showing a larger descriptive effect than AI teaching (g = 0.69). → Mark Feng Teng 2026 q3 i? When controlling for risk of bias and sample size, the traditional-classroom coefficient became statistically non-significant (β = 0.74, p = .230), so the subgroup difference is not attributed to pedagogical format itself. → Mark Feng Teng 2026 (2)

Evidence¶

Mark Feng Teng 2026¶

Mark Feng Teng. (2026). Effectiveness of GenAI in enhancing writing performance: a meta-analysis. Asian-Pacific Journal of Second and Foreign Language Education. https://doi.org/10.1186/s40862-026-00444-2

q3 · i? · quant-synthesis · r2

Subgroup moderator analysis in the expanded main synthesis (k = 11). Traditional classroom studies (k = 3, g = 1.19, 95% CI [0.73, 1.64]) showed a larger pooled effect than AI teaching studies (k = 8, g = 0.69, 95% CI [0.41, 0.96]); QB = 27.26, p < .001.

"the subgroup analysis initially indicated that the instructional model was a significant source of between-study heterogeneity ( QB = 27.26, p <. 001), with traditional classrooms associated with a larger descriptive effect size (g=1.19) than AI-teaching contexts (g=.69)."

Mark Feng Teng 2026 (2)¶

Mark Feng Teng. (2026). Effectiveness of GenAI in enhancing writing performance: a meta-analysis. Asian-Pacific Journal of Second and Foreign Language Education. https://doi.org/10.1186/s40862-026-00444-2

q3 · i? · quant-synthesis · r3

Multivariate robust meta-regression (HC3) controlling for risk-of-bias classification and sample size. The traditional-classroom coefficient was "statistically non-significant (β = 0.74, p = .230)", and the article attributes the initial subgroup difference to confounding rather than pedagogical format.

"the coefficient for the traditional classroom model became statistically non-significant (β = 0.74, p = .230). This suggests that the initial subgroup differences may be partially confounded by study-level methodological quality or small-study effects, rather than the pedagogi -cal format itself."

Discussion¶

Related Claims¶

• Study setting significantly moderates GenAI productivity effects: laboratory experiments show large significant gains (g = 0.73) while enterprise (g = 0.19) and open-source (g = 0.01) effects are substantially smaller and non-significant. — related
• Risk-of-bias classification is the only robust moderator of GenAI writing effects, with unrated studies showing significantly larger effects than high-risk studies — related
• Instructional method does not significantly moderate Gen-AI's effect on HOT, though project-based learning shows the largest subgroup effect (ES = 0.717) and lecture-based instruction a small non-significant effect (ES = 0.396) — related
• The effect of ChatGPT on learning performance is moderated by type of course, learning model, and duration, with problem-based learning and 4–8-week interventions showing the largest effects — related
• Personalized learning models combining classroom use with student-centered pedagogies (Model C) exert larger effects than fully automated platforms (Model A) — related