Does low confidence drive reliance on AI, or does reliance erode confidence? A four-wave RI-CLPM study of chinese language and literature undergraduates

Generative AI tools are now routinely used by Chinese Language and Literature majors for tasks once considered the province of individual literary craft: drafting verse, shaping narrative prose, revising fiction. Two competing accounts describe what this means for students' confidence in their own creative writing. A cognitive-offloading account holds that repeated reliance on AI displaces the mastery experiences through which writing self-efficacy is built, so reliance erodes confidence over time. A compensatory-use account holds the reverse: students whose confidence is already low turn to AI to avoid the discomfort of unassisted composition, so low confidence drives reliance. Cross-sectional and standard cross-lagged designs cannot separate these accounts because they confound stable between-person differences with genuine within-person change. We addressed this using a random-intercept cross-lagged panel model (RI-CLPM) across four waves spanning one academic year among Chinese Language and Literature undergraduates (N = 390 recruited; N = 387 analyzed under full-information maximum likelihood, of whom 298 supplied complete four-wave data). Generative AI writing reliance (GAWR) and literary creative writing self-efficacy (LCWSE) were assessed at each wave. The trait-level association between the two constructs was negative and could not be distinguished from zero (r = -.342, Δχ²(1) = 2.26, p = .133). At the within-person level, declines in self-efficacy predicted subsequent increases in reliance (β = -.276, p < .001). The reverse path, from reliance to subsequent self-efficacy, was smaller and fell short of the conventional threshold (β = -.085, p = .072). We state explicitly that this second result is not interpretable as evidence against erosion: a Monte Carlo power analysis put this design's power to detect an effect of that magnitude at 53.2%, close to a coin flip, so the erosion pathway is undetermined in these data and this null supplies no basis for concluding that erosion does not operate. A direct comparison of the two cross-lagged paths favored compensation (Δχ²(1) = 8.83, p = .003; bootstrap 95% CI for the difference [.020, .218] across 5,000 resamples), with the qualification that this comparison sets an adequately powered estimate against an underpowered one and therefore describes what this design was able to detect. Compensation is the pathway this study establishes; whether reliance also erodes confidence over a longer horizon is an open question. We discuss implications for writing pedagogy in Chinese Language and Literature programs and for how underpowered null results should be interpreted in AI-in-education panel research.