When does GenAI work for L2 learning? A three‐level meta‐analysis

This three‐level meta‐analysis examined when generative AI (GenAI) works for second language (L2) learning, synthesizing 67 effect sizes from 48 controlled studies ( N = 5196) published between 2023 and 2026, with a focus on GenAI's text generation and conversational interaction capabilities. GenAI demostrated a large overall effect on measured L2 learning outcomes ( g = 0.76, 95% CI [0.50, 1.02]); after removal of influential outliers, the effect remained significant but was smaller ( g = 0.57). Effects varied by skill domain: speaking showed the largest full‐sample effect ( g = 1.48), which declined substantially after outlier adjustment ( g = 0.65), while vocabulary ( g = 0.66) and writing ( g = 0.61) remained moderate and reading was non‐significant. Regarding integration design, GenAI functioning as a feedback and assessment provider yielded the largest effects (full‐sample g = 1.15; adjusted g = 0.83), and interaction configurations with higher learner agency (Student‐Led, g = 0.92; Co‐Creative, g = 0.82) outperformed AI‐directed ones (AI‐Led, g = 0.34). Teacher role did not significantly moderate effects in the primary analysis; an observer‐teacher advantage ( g = 0.99) emerged only after outlier removal and should be considered exploratory. These findings suggest that GenAI effectiveness depends on instructional configuration, including the functional role assigned to the tool, the learner–AI interaction mode and the nature of teacher involvement, rather than representing an inherent advantage of the technology itself.