A helping hand or a dominant partner? Individual perceptions of GenAI reliance and human agency in collaborative learning

Synthesis: Wu and Lu examine whether students' self-reported reliance on generative AI predicts a subsequent loss of perceived human agency in AI-supported collaborative learning, using three waves of data from 342 undergraduates working in 85 fixed groups in an AI-supported collaborative writing course. They estimated both a traditional cross-lagged panel model and a random intercept cross-lagged panel model to separate observed-level associations from within-person temporal relations after accounting for stable between-person differences. In the RI-CLPM, higher-than-usual reliance was associated with lower subsequent agency, whereas agency-to-reliance paths were weaker and not statistically supported. Equality-constraint tests were consistent with directional asymmetry, though confidence intervals and Monte Carlo sensitivity analyses showed that small reverse effects remain possible. Backend logs confirmed that self-reported reliance corresponded to AI-use intensity, but log indicators could not distinguish strategic consultation from Cognitive Offloading or deference.

Key Findings

1. Reliance predicts declining agency, not the reverse. In the random-intercept model, higher-than-usual reliance predicted lower subsequent perceived human agency, while agency-to-reliance paths were weaker and not statistically supported.
2. Directional asymmetry, but with residual uncertainty. Equality-constraint tests were consistent with reliance having the stronger longitudinal path, while 95% confidence intervals and Monte Carlo sensitivity analyses showed small reverse effects remain possible.
3. Reliance tracks AI-use intensity. Backend logs showed self-reported reliance corresponded to how intensely students used AI, but the logs could not distinguish strategic consultation from cognitive offloading or deference.
4. A three-wave panel design. Three-wave data came from 342 undergraduates in 85 fixed groups in a collaborative writing course, analyzed with both CLPM and RI-CLPM.
5. The practical lever is visibility. The authors conclude that AI-supported collaborative tasks should keep students' responsibility for interpretation, judgment, and authorship visible throughout the learning process.

Reliance, agency, and the look within person

The study responds to a recurring concern in AI-supported learning: that system-generated suggestions may reduce opportunities for judgment, negotiation, and ownership. Traditional cross-lagged panel models, however, conflate stable between-person differences with within-person temporal associations, so the authors estimated a random intercept cross-lagged panel model that separates the two. The RI-CLPM result — higher-than-usual reliance predicting lower subsequent agency — provides stronger evidence for the reliance-to-agency direction than for the reverse, while acknowledging that small reverse effects cannot be ruled out.

The design situates the question in authentic practice: an AI-supported collaborative writing course in which students worked together and used GenAI for Feedback, explanation, and drafting support. The use of 85 fixed groups and three waves allowed within-person temporal relations to be examined after between-person differences were accounted for, a distinction that matters because students who chronically rely on AI differ from students whose reliance changes wave to wave.

Using logs without equating activity with over-reliance

A distinctive finding concerns how AI-use intensity is interpreted. Backend logs confirmed that students' self-reported reliance corresponded to how intensely they used AI. Yet the authors caution that high log-based AI-use intensity is not the same as over-reliance, because log indicators cannot distinguish strategic consultation from cognitive offloading or deference. The implication for learning analytics is direct: dashboards should not equate high AI-use intensity with over-reliance unless log data are interpreted alongside discourse, reflection, revision, or decision-making evidence. This matters for how the field measures dependence, since the same behavior can signal thoughtful use or outsourcing depending on the context.

What this means for practice

• Instructors. Ask groups to form an initial position before invoking AI and to explain why AI suggestions are accepted, revised, or rejected, so responsibility for interpretation and authorship stays visible.
• Instructors. Design GenAI tools to support agency by offering criticism and prompting reflection while leaving decisions about task direction to students, rather than delivering ready answers.
• Learning analytics designers. Do not equate high AI-use intensity with over-reliance; interpret log data alongside evidence of discourse, reflection, and revision before flagging dependence.
• Researchers. Use random-intercept cross-lagged designs when studying reliance and agency, because traditional CLPMs conflate stable between-person differences with within-person change.

Limitations

• The outcome is perceived human agency, not observed agency or measured learning outcomes.
• Log-based AI-use data could corroborate reliance intensity but could not distinguish strategic consultation from cognitive offloading or deference.
• The study is a single collaborative writing course at one institution, so generalization to other tasks and contexts is limited.
• Small reverse effects (agency-to-reliance) remain possible, as the confidence intervals and Monte Carlo sensitivity analyses show.

Citation

Wu, Y., & Lu, X. (2026). A helping hand or a dominant partner? Individual perceptions of GenAI reliance and human agency in collaborative learning. British Journal of Educational Technology. Advance online publication.