Language learners’ trust in AI was linked to reliance and resistance, with self-direction shaping reported use

Trust in generative AI is not simply a quality to maximize. A learner might trust a tool enough to explore its suggestions, or trust it so readily that checking stops. Feifei Wang and colleagues investigated this distinction in language learning through an explanatory mixed-methods study. Their findings connect trust with reliance, resistance, and reported use, while showing how students describe different ways of negotiating those relationships. The work addresses learner behavior and interpretation rather than testing whether AI improves language proficiency.

The quantitative phase recruited Chinese university students who had already used generative AI for language learning. An online survey collected between June and August 2024 produced 682 valid responses. The researchers used established measures adapted and translated for the context, then tested a structural equation model involving trust, reliance, resistance, behavioral intention, self-directed learning, and use. A second phase interviewed 40 purposively selected participants to explore the survey relationships and capture experiences that a path diagram could not fully describe.

In the statistical model, greater trust was associated with stronger intention to use AI and greater reliance, while resistance showed an opposing relationship with intention. Self-directed learning moderated the association between intention and reported use. These relationships were consistent with the authors’ proposed framework, but consistency does not establish temporal order. Students were not assigned different levels of trust or self-direction, and the measures came from the same survey occasion. A fitted model cannot by itself prove that increasing trust will produce responsible or effective use.

The interviews give the statistical results a more practical meaning. Some learners described accepting suggestions because they assumed the system knew more than they did. Others checked whether an answer addressed their question, revised prompts, or retained responsibility for the final text. Resistance also took different forms, including preference for familiar learning methods and more active skepticism following disappointing experiences. Learners linked their decisions to perceived accuracy, academic-integrity expectations, data concerns, and the clarity of support from teachers or institutions.

These accounts suggest that reliance is a judgment made in context, not a single stable trait. A student may accept a grammar suggestion while rejecting an interpretation of an argument. The researchers describe appropriate, excessive, and insufficient reliance, alongside disuse and different kinds of inappropriate use. Those categories are conceptual interpretations of participants’ accounts. They should not be read as externally verified classifications of every learner’s actual behavior, or as evidence that one pattern caused a particular language-learning outcome.

The authors explicitly identify the cross-sectional design, self-report bias, and the concentration on language learners in Chinese higher education as limitations. The survey’s use measure was not a platform-log observation, and no independent language assessment established educational benefit. The interview sample deepens understanding but does not remove those boundaries. Future longitudinal or experimental work would be needed to test how trust develops, how it responds to experience, and whether changes in self-directed practice alter learning.

For AIEDHK, the implication is to support calibrated judgments rather than blanket confidence or blanket rejection. Educators could ask learners to state the purpose of an AI request, check a disputed response against a reliable resource, and explain which decisions they retained. Clear course expectations would make those judgments easier to discuss. These are proposed classroom applications of the study’s framework, with success still requiring observation of both actual behavior and independent language performance.