Low Writing Confidence Drives Students Toward AI, Four-Wave Study Finds

When Chinese Language and Literature undergraduates lose confidence in their own creative writing, they tend to reach for generative AI tools, according to a new longitudinal study. The research, which tracked students across four waves spanning a full academic year, set out to disentangle two competing explanations for why aspiring writers lean on artificial intelligence when composing verse, narrative prose, and fiction. The first explanation, known as the cognitive-offloading account, holds that repeated reliance on AI displaces the mastery experiences through which writing self-efficacy is built.

In this view, students who delegate drafting and revising to machines gradually lose faith in their own abilities because they no longer practice unassisted composition. The second explanation, the compensatory-use account, reverses the arrow: students whose confidence is already low turn to AI precisely to avoid the discomfort of writing on their own. Previous research could not separate these accounts cleanly.

Cross-sectional surveys capture a single moment in time, and standard cross-lagged panel designs confound stable differences between people with genuine change within each person. A student who both doubts herself and uses AI heavily may simply have been that way all along, rather than one factor driving the other. To break that deadlock, the researchers employed a random-intercept cross-lagged panel model, or RI-CLPM, a statistical approach that partitions each construct into a stable trait-level component and a time-varying within-person component.

Only the within-person part is examined for reciprocal dynamics, allowing the study to ask whether fluctuations in one variable predict later fluctuations in the other within the same individual. The sample comprised 390 recruited undergraduates, of whom 387 were analyzed under full-information maximum likelihood; 298 provided complete data at all four waves. At each wave, the team measured generative AI writing reliance and literary creative writing self-efficacy.

At the trait level, the association between the two constructs was negative but statistically indistinguishable from zero (r = −.342, Δχ²(1) = 2.26, p = .133). The within-person results told a clearer story. Declines in self-efficacy predicted subsequent increases in reliance on AI, with a path coefficient of β = −.276, p < .001.

The reverse path—from reliance to later self-efficacy—was smaller (β = −.085) and fell short of the conventional significance threshold at p = .072. Crucially, the authors state explicitly that this second result cannot be interpreted as evidence against the erosion hypothesis. A Monte Carlo power analysis showed the design had only a 53.2% chance of detecting an effect of that magnitude—barely better than a coin flip.

The erosion pathway therefore remains undetermined in these data, and the near-null supplies no basis for concluding that erosion does not operate. A direct comparison of the two cross-lagged paths favored the compensation account (Δχ²(1) = 8.83, p = .003; bootstrap 95% CI for the difference [.020, .218] across 5,000 resamples). The researchers add an important qualification: this comparison pits an adequately powered estimate against an underpowered one, so it describes what this particular design was able to detect rather than settling the question definitively.

The findings carry implications for writing pedagogy in Chinese Language and Literature programs, where generative AI tools are now routinely used for tasks once considered the province of individual literary craft. The study also offers a methodological lesson for AI-in-education research: underpowered null results should be reported with their power limitations stated openly, as this team did, rather than treated as evidence of absence. Whether reliance on AI also erodes creative writing confidence over a longer horizon remains an open question for future work.

Subject of Research: Longitudinal reciprocal relationships between generative AI writing reliance and literary creative writing self-efficacy in Chinese Language and Literature undergraduates

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

Article References: Du, X., & Yu, S. (2026). Does low confidence drive reliance on AI, or does reliance erode confidence? A four-wave RI-CLPM study of chinese language and literature undergraduates. PLOS One, 21(10), e0361489. https://doi.org/10.1371/journal.pone.0361489

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0361489

Keywords: generative AI, writing self-efficacy, RI-CLPM, longitudinal study, cognitive offloading, compensatory use, creative writing, higher education, Chinese Language and Literature, AI in education, panel study, statistical power