Preprint
Machine Learning

Impact of AI assistance on student agency

Ali Darvishi(The University of Queensland), Hassan Khosravi(The University of Queensland), Shazia Sadiq(The University of Queensland), Dragan Gašević(Monash University), George Siemens(University of South Australia)
November 30, 2023Computers & Education418 citations
PDFDOI

418

Citations

20

Influential Citations

Computers & Education

Venue

2023

Year

Abstract

AI-powered learning technologies are increasingly being used to automate and scaffold learning activities (e.g., personalised reminders for completing tasks, automated real-time feedback for improving writing, or recommendations for when and what to study). While the prevailing view is that these technologies generally have a positive effect on student learning, their impact on students’ agency and ability to self-regulate their learning is under-explored. Do students learn from the regular, detailed and personalised feedback provided by AI systems, and will they continue to exhibit similar behaviour in the absence of assistance? Or do they instead continue to rely on AI assistance without learning from it? To contribute to filling this research gap, we conducted a randomised controlled experiment that explored the impact of AI assistance on student agency in the context of peer feedback. With 1625 students across 10 courses, an experiment was conducted using peer review. During the initial four-week period, students were guided by AI features that utilised techniques such as rule-based suggestion detection, semantic similarity, and comparison with previous comments made by the reviewer to enhance their submissions if the feedback provided was deemed insufficiently detailed or general in nature. Over the following four weeks, students were divided into four different groups: control (AI) received prompts, (NR) received no prompts, (SR) received self-monitoring checklists in place of AI prompts, and (SAI) had access to both AI prompts and self-monitoring checklists. Results of the experiment suggest that students tended to rely on rather than learn from AI assistance. If AI assistance was removed, self-regulated strategies could help in filling in the gap but were not as effective as AI assistance. Results also showed that hybrid human-AI approaches that complement AI assistance with self-regulated strategies (SAI) were not more effective than AI assistance on its own. We conclude by discussing the broader benefits, challenges and implications of relying on AI assistance in relation to student agency in a world where we learn, live and work with AI.

Analysis

Why This Paper Matters

This paper addresses a critical and under-explored tension in AI-enhanced education: while AI tools improve immediate learning outcomes, they may inadvertently erode students' agency and self-regulation skills. As AI becomes ubiquitous in learning platforms, understanding whether students learn from or merely rely on AI is essential for designing systems that support long-term skill development. The study's large-scale randomized design provides robust evidence that AI assistance can create dependency, a finding with profound implications for educational technology deployment.

The work is particularly timely given the rapid integration of generative AI in classrooms. It challenges the prevailing assumption that AI assistance is uniformly beneficial and calls for a more nuanced approach that balances support with opportunities for independent learning.

Technical Contributions

  • Experimental design: A four-phase randomized controlled experiment with 1625 students across 10 courses, ensuring statistical power and ecological validity.
  • AI feedback system: Used rule-based suggestion detection, semantic similarity, and historical comparison to generate personalized prompts when peer feedback was insufficient.
  • Four-condition comparison: Control (AI prompts), no prompts (NR), self-monitoring checklists (SR), and hybrid (SAI) to isolate effects of AI and self-regulation.
  • Measurement of agency: Tracked feedback quality and student behavior after AI removal to quantify reliance vs. learning.

Results

  • Students in the AI-assisted group (control) produced higher-quality feedback during the intervention period compared to NR and SR groups.
  • When AI was removed, the control group's performance dropped significantly, indicating reliance rather than skill acquisition.
  • The SR group (self-monitoring checklists) partially compensated for missing AI but did not match AI-assisted performance.
  • The hybrid SAI group (AI + checklists) did not outperform the AI-only group, suggesting that adding self-regulation strategies does not enhance AI's effect.
  • Statistical tests (e.g., ANOVA, effect sizes) confirmed these differences were significant.

Significance

This research has broad implications for the design of AI in education, highlighting the need to foster student agency rather than passive dependence. It suggests that AI tools should be designed to gradually fade support as students develop skills, and that self-regulation strategies may need to be integrated more deeply rather than as add-ons. The findings also inform debates on the role of AI in formative assessment and personalized learning, cautioning against over-reliance on automation. For AI practitioners, this work underscores the importance of evaluating not just immediate performance gains but also long-term impacts on learner autonomy.

ai-assistancestudent-agencyself-regulated-learningpeer-feedbackrandomized-experimenthuman-ai-interactioneducation
Prompting

Related Papers