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

Synthesis: Wang and Li ran two studies asking whether AI-assisted speaking practice is associated with readiness to speak with humans rather than with AI. Study 1 was a cross-sectional survey of 708 English learners; Study 2 was a 2 × 2 scenario comparison (N = 726) in which participants imagined preparing with an AI voice assistant or smartphone recording for a human group discussion that was either formally graded or ungraded. Greater AI-assisted speaking behavior was associated with higher human-directed willingness to communicate (β = 0.384) and lower foreign-language speaking anxiety (β = −0.416), with a positive statistical indirect association through anxiety (0.127, 95% CI [0.079, 0.184]). AI-voice-assistant scenarios showed lower anticipated anxiety and higher anticipated willingness, but the tool contrast bundled AI identity with dialogue, responsiveness and feedback, and neither study observed actual human speaking.

Key Findings

• AI-assisted speaking behavior tracked human-directed readiness in the survey. With N = 708, greater AISB was positively associated with HWTC (B = 0.359, β = 0.384, p < 0.001, R2 = 0.148) and negatively associated with FLSA (B = −0.404, β = −0.416, p < 0.001).
• Anxiety carried a positive indirect association. FLSA remained negatively associated with HWTC when AISB was controlled (B = −0.313, β = −0.325, p < 0.001), and the statistical indirect association from AISB to HWTC through FLSA was 0.127, 95% CI [0.079, 0.184].
• Speaking self-efficacy changed the anxiety–readiness link, with an upper-tail crossover. The FLSA × SSE interaction was significant (B = 0.281, p < 0.001): the FLSA–HWTC slope was −0.371 at low SSE and +0.137 at high SSE, and the conditional indirect association fell from 0.142 at low SSE to 0.044 at mean SSE and reversed to −0.054 at high SSE.
• The crossover was concentrated in a small upper tail. Johnson–Neyman probing placed transition points at SSE ≈ 3.88 and ≈ 4.45, with only 55 participants (7.8%) above the upper threshold.
• In the scenarios, the AI voice assistant was associated with lower anticipated anxiety and higher anticipated willingness. Tool effects were F(1,722) = 171.07, ηp2 = 0.192 for FLSA and F(1,722) = 116.97, ηp2 = 0.139 for HWTC, while formal evaluation moved both outcomes the other way (ηp2 = 0.102 and 0.061).
• The tool gap widened under evaluation. For FLSA the AI–recording difference was ∆M = −1.10 under evaluation against ∆M = −0.50 without it; for HWTC it was ∆M = +1.05 (d = 1.35) against ∆M = +0.30 (d = 0.34). Both interactions were significant (ηp2 = 0.032 and 0.047).
• Willingness and speaking-initiation tendency diverged. Group B (AI × no evaluation) had the lowest FLSA (M = 2.70) and highest HWTC (M = 3.60); Group C (recording × evaluation) had the highest FLSA (M = 4.05) and lowest HWTC (M = 2.50). Yet the AI–recording difference in speaking-initiation tendency was larger without evaluation (d = 0.58) than under it (d = 0.31).

What the two evidence types each support

Study 1 is a one-time questionnaire: recent AI-assisted spoken English behavior (AISB) over the preceding four weeks, foreign-language speaking anxiety (FLSA), speaking self-efficacy (SSE) and human-directed willingness to communicate (HWTC), each on five contextualized items (α = 0.847, 0.834 and 0.833). The three-factor model fit well (χ2 (87) = 115.31, CFI = 0.993, RMSEA = 0.021, SRMR = 0.024), HTMT values ran 0.485 to 0.511, but HWTC's AVE was 0.499 — borderline against the conventional 0.50 Benchmark, which the authors flag. All of these associations rest on self-report, and the design cannot establish temporal order. Demographic adjustment left the pattern essentially unchanged (AISB→HWTC B = 0.357, β = 0.382; FLSA→HWTC B = −0.312, β = −0.324; adjusted indirect association 0.126, 95% CI [0.078, 0.182]), and an unmeasured latent method factor sensitivity model produced materially stable estimates (AISB→FLSA moved from −0.429 to −0.428, FLSA→HWTC from −0.496 to −0.491, the indirect estimate from 0.213 to 0.210), with modest method loadings (range −0.265 to 0.212, mean absolute 0.099).

Study 2 does not measure behavior either. Participants read that they would join a five-minute English discussion with two other students after ten minutes of preparation, then reported anticipated FLSA and HWTC. The manipulation checks registered the contrasts (tool interactivity 3.90 versus 2.50, t(707.2) = 23.32, d = 1.74; evaluation pressure 3.90 versus 2.50, t(697.3) = 23.89, d = 1.78), and covariate-adjusted ANCOVAs kept every scenario contrast significant: for FLSA, tool F(1,716) = 204.90, ηp2 = 0.223, evaluation F(1,716) = 95.26, ηp2 = 0.117, interaction F(1,716) = 31.30, ηp2 = 0.042; for HWTC, tool F(1,716) = 140.95, ηp2 = 0.164, evaluation F(1,716) = 57.47, ηp2 = 0.074, interaction F(1,716) = 39.01, ηp2 = 0.052. Because assignment used separate survey links rather than individual randomization, and no preparation was actually delivered, the authors call these controlled scenario-condition comparisons, not treatment effects. Scenario-realism means ran from approximately 3.15 to 3.30, which the authors read as imaginable but not equivalent to a real classroom.

What this means for practice

• Instructors. Treat reduced anxiety as a bridge rather than an outcome. The survey's positive indirect association (0.127) runs through FLSA, and nothing here shows that lower anxiety produces better human speaking.
• Instructors. Stage the evaluative load: ungraded interactive rehearsal first, then human peer interaction, then formal grading. In Study 2 the AI × no-evaluation cell still had the lowest FLSA and highest HWTC, so raising evaluative threat is not what the pattern recommends.
• Curriculum designers. Do not let AI rehearsal substitute for human interaction. Both studies stop at readiness and anticipated affect, and the authors call the transfer pathway a pedagogical proposal to be tested directly.
• Assessment designers. Where high-stakes speaking assessment is unavoidable, use voice tools as rehearsal and base grades on human performance without AI, rather than assuming AI practice replaces valid evaluation anxiety.
• Platform designers. Track process data — practice turns, follow-up questions generated, feedback taken up, re-recordings — so usage intensity can be separated from supportiveness and from human performance.

Limitations

• Study 1 is cross-sectional and entirely self-report, with no temporal separation, independent data source or marker variable; the ULMC analysis reduced but did not eliminate common-method concerns, and no causal direction is established.
• The shortened FLSA, SSE and HWTC forms were not validated by an independent expert panel or a separate pilot, and HWTC's AVE of 0.499 sits marginally below the 0.50 benchmark.
• Study 2 rests on one-off imagined scenarios assigned through separate survey links rather than individual randomization, so the contrasts are not treatment effects, and the manipulation checks may partly reflect demand characteristics.
• Neither study observed human speaking or measured objective proficiency, and the AI-voice-assistant condition bundled AI identity with dialogic interaction, responsiveness and feedback, so the difference cannot be attributed to AI alone.

Citation

Wang, R., & Li, J. (2026). AI-Assisted Speaking Practice and Human-Directed Communicative Readiness: Survey and Scenario Evidence on Anxiety, Self-Efficacy, and Evaluation Pressure. Behavioral Sciences, 16(10), 1770.