Background: Artificial intelligence (AI) is increasingly used in second-language (L2) instruction, yet the size and consistency of effects in writing, speaking, listening, and reading remain uncertain. Methods: We conducted a meta-analysis of randomized controlled trials comparing AI-assisted with non-AI instruction. Primary outcomes were standardized post-test performance in the four domains; self-reported outcomes (engagement, flow, anxiety, self-regulation) were narratively synthesized. Effect sizes were computed as standardized mean differences (Hedges’ g) and pooled with random-effects models. Between-study heterogeneity was quantified (tau-squared, I-squared) and examined with influence diagnostics and leave one-out sensitivity analyses. Results: Four trials were eligible for quantitative synthesis. Across all endpoints, effects favored AI-assisted instruction. The pooled estimate indicated a moderate benefit, with heterogeneity largely explained by domain and program duration. Listening showed the largest gains; speaking and writing showed moderate gains; reading showed smaller statistically reliable improvements. Findings were stable to alternative data-handling choices. One reading trial using the Gray Silent Reading Test reported g = 0.34 (95% CI 0.21-0.48). Narrative evidence suggested higher engagement and lower anxiety in AI-supported conditions. Conclusions: AI-assisted L2 instruction improves achievement with moderate effects and domain-specific variation. Benefits appear greatest for listening and speaking, within feedback-intensive, deliberate-practice workflows.