This year-long quasi-experimental study examined whether sub-second, generative-AI feedback loops can engineer durable improvements in learners emotion-regulation capacity and, by extension, accelerate spoken foreign-language development. Two hundred and fourteen Grade-9 Chinese EFL learners (107 experimental, 107 control) completed 128 teaching hours over thirty-two weeks. In the experimental arm, ChatGPT-4o streamed affect-sensitive feedback to individual tablets with a median detection-to-response latency of 0.68 s, whereas control classes received conventional end-of-task comments. Forty-one multimodal cues, eye-fixation entropy, galvanic-skin-response peaks, voice-stress jitter, lexical surprisal, disfluency density and more, were synchronised at 30 Hz on an IEEE-1588 backbone and fused by a dynamic Bayesian network into a continuous Emotion Regulation Index (ERI). AI-triggered micro-interventions occurred 14 410 times (mean duration = 9.3 s) and were logged for lag analysis. Latent growth-curve modelling showed a robust positive ERI slope in the AI cohort ( = 0.083 0.011, p < 0.001) and a marginal decline among controls ( = 0.019 0.014, p = 0.071). Year-end Cambridge B1 speaking scores rose 1.34 SD for AI learners versus 0.57 SD for controls (t = 9.27, df = 212, p < 0.0001). Results indicate that finely tuned generative-AI feedback stabilises classroom affect, shortens self-regulation latency, and delivers substantial proficiency dividends.
Zhao et al. (2025) studied this question.