AI-Assisted Speaking Practice and Human-Directed Communicative Readiness: Survey and Scenario Evidence on Anxiety, Self-Efficacy, and Evaluation Pressure

Generative artificial intelligence (GenAI) voice tools create low-risk opportunities for second-language (L2) speaking practice, but it remains unclear whether AI-assisted speaking practice is associated with human-directed communicative readiness and how this relationship varies across learners and evaluation contexts. Study 1 analysed cross-sectional survey data from 708 English learners. Greater AI-assisted spoken English behaviour (AISB) was associated with higher human-directed willingness to communicate (HWTC; β = 0.384) and lower foreign-language speaking anxiety (FLSA; β = −0.416); FLSA was negatively associated with HWTC when AISB was controlled (β = −0.325), yielding a positive statistical indirect association (0.127, 95% CI [0.079, 0.184]). Speaking self-efficacy (SSE) moderated the FLSA–HWTC association, including an unexpected upper-tail crossover. Demographic adjustment left the substantive pattern essentially unchanged, and an unmeasured latent method factor (ULMC) sensitivity analysis showed materially stable focal factor-score associations after accounting for a general method factor. Study 2 used a 2 × 2 nonrandom scenario comparison (N = 726), in which participants imagined preparation with an AI voice assistant versus smartphone recording under formal evaluation versus no evaluation. AI-voice-assistant scenarios showed lower anticipated FLSA and higher anticipated HWTC, whereas formal evaluation showed the opposite pattern; tool × evaluation interactions were significant for FLSA (ηp2 = 0.032) and HWTC (ηp2 = 0.047). Speaking-initiation tendency showed a different interaction pattern, indicating that HWTC and speaking-initiation tendency are not identical. The findings support a competence–affect–context account of AI-assisted speaking practice and human-directed communicative readiness, but they do not establish AI-specific causal effects or transfer to actual human speaking.