The Inference Ceiling of Generative AI Effects in English Language Learning

Claims that generative artificial intelligence (GenAI) improves language learning depend on the counterfactual, the assessment condition, and the time horizon. We conducted a convergent segregated mixed-methods evidence audit of GenAI-supported English-language learning and introduced an inference-ceiling framework that matches each claim to its minimum design requirements. The verified database consolidated 676 source-study rows into 530 canonical records; 205 studies met direct scope, including 29 controlled studies (N = 3,231), 78 variance-complete controlled effects, 545 direct-scope mechanism findings from 96 studies, and 51 direct-scope mixed-methods bridges from 46 studies. Random-effects models used restricted maximum likelihood, Hartung-Knapp confidence intervals, and prediction intervals. Objective language performance favored GenAI-supported packages over usual practice (k = 10, g = 1.23, 95% CI [0.44, 2.03], prediction interval [-1.00, 3.47]) but was smaller and imprecise against established active alternatives (k = 7, g = 0.71, 95% CI [-0.06, 1.48], prediction interval [-1.41, 2.83]). A post hoc active-minus-usual coefficient was -0.51 (95% CI [-1.54, 0.53], p = .316), so descriptive attenuation was not statistically distinguishable from zero. After full-text verification, assessment access remained unreported in the source for 74 of 78 controlled effects; only one explicitly used tool-withdrawn assessment, and no variance-complete effect identified GenAI's incremental contribution within a common instructional base. Study-clustered mechanism sensitivity showed that negative or boundary evidence appeared in 76.2% of studies contributing to the offloading/integrity family. Current evidence supports context-dependent package benefits, not a stable estimate of GenAI-specific or independently retained learning.