This research examines the impact of feedback timing on metacognition in the context of AI-mediated language learning. The study addresses a gap in the literature, since most previous works have primarily focused on human-to-human interactions, overlooking the influence of AI-mediated feedback on metacognition. Thus, the objective of this research is to explore how feedback timing affects emotional and cognitive states (namely, valence, arousal, and cognitive load) and how these states mediate the relationship between feedback timing and metacognitive judgment. A between-subjects experiment involved 30 adult English speakers learning French at the A2 French language proficiency level. Participants were randomly assigned to an experimental condition involving either immediate or delayed feedback during a reading-aloud task in French, where an AI tutor provides pronunciation feedback. Data collection involved both self-reported and physiological measure. The results indicate that feedback timing did not have an impact on valence, arousal or cognitive load, nor in metacognitive judgment accuracy. However, emotional valence was positively associated with higher metacognitive accuracy in comprehensibility, while increased cognitive load was associated with improved accuracy in metacognitive judgments for accentedness. This implies that, although feedback timing may not have a direct effect on metacognitive outcomes, emotional valence and cognitive load were found to have a significant effect on improving metacognitive accuracy.