How Students Use AI Feedback Matters: Experimental Evidence on Physics Achievement and Autonomy

“When learners perceive autonomy in their educational environment, they are more likely to exhibit intrinsic motivation, leading to deeper engagement and improved learning outcomes (Deci & Ryan, 2000).”

“Instead, it offered fundamental conceptual principles, contextual analysis, and directional suggestions for problem-solving attempts, thereby providing optimal scaffolding for independent learning (Kirschner et al., 2006).”

“…in the domain of STEM knowledge and skill learning, where direct measurements of students’ self-reported autonomy have similarly indicated, consistent with Darvishi et al. (2024), that AI-provided learning scaffolds may enhance academic performance while concurrently diminishing learner autonomy.”

“…Zimmerman’s cyclical model, Pintrich’s self-regulation framework, and Winne and Hadwin’s information processing model, emphasize the dynamic nature of SRL, highlighting its role in goal setting, self-monitoring, and feedback-based adjustments (Zimmerman, 2011; Pintrich, 2000; Winne & Hadwin, 1998).”

“These principles were explicitly integrated into the workflow’s prompts, and chain-of-thought examples (Wei et al., 2022) were provided for agent training and reference.”

“Similarly, Anderson et al. (1995) developed Cognitive Tutors based on the ACT theory, which provided students with step-by-step hints during problem-solving tasks.”

“…Zimmerman’s cyclical model, Pintrich’s self-regulation framework, and Winne and Hadwin’s information processing model, emphasize the dynamic nature of SRL, highlighting its role in goal setting, self-monitoring, and feedback-based adjustments (Zimmerman, 2011; Pintrich, 2000; Winne & Hadwin, 1998).”

“Therefore, whether considering the potential benefits or risks, it is crucial to adopt a rational and clearly bounded perspective when assessing the impact of GAI on education and learning.”

“…these models share a common focus on the interplay between cognition, motivation, and behavioral control as key components of SRL. Self-regulation is regarded as a crucial ability fostering both short-term academic success and long-term developmental achievements (McClelland & Cameron, 2012).”

“(Aleven et al., 2003).”

“In a flipped classroom context, Huang et al. (2023) applied AI-enabled personalized video recommendations for 43 college students studying systems programming and compared their outcomes with 59 control participants; results indicated that personalized recommendations significantly enhanced…”

“Such evidence resonates with concerns raised by Stadler et al. (2024) and Darvishi et al. (2024), who caution that GAI-powered cognitive scaffolds diminish deep thinking and agency.”

“Moreover, existing studies often adopt a limited practical approach, primarily facilitating direct interactions between students and Chat-GPT without systematically exploring well designed learning systems underpinned by pedagogical frameworks and empowered by GAI (Deng et al., 2025).”

“In the context of widespread adoption of GAI for on-demand help in education, research into the educational impacts and underlying mechanisms of GAI-based intelligent learning systems has become critically important (Yan et al., 2024).”

“Renkl (2002) conducted an experimental study with 60 participants to compare the effects of instructional explanations versus prompting self-explanations, finding that the latter led to better learning outcomes.”

“Driven by the rapid development of generative artificial intelligence (GAI), numerous technological innovations have emerged to enrich K–12 education, particularly in STEM domains where robust logical reasoning and sophisticated pedagogical design are essential (García-Martínez et al., 2023).”

“These contemporary understanding of autonomy aligns with the philosophical perspective wherein moral agency arises from individuals acting according to principles they have rationally chosen for themselves (Kant, 1992).”

“Hsu et al. (2013) found that recommending reading materials aligned with students’ proficiency significantly reduced cognitive load.”

“Such evidence resonates with concerns raised by Stadler et al. (2024) and Darvishi et al. (2024), who caution that GAI-powered cognitive scaffolds diminish deep thinking and agency.”

“The usage patterns reflect pedagogy and rules of student learning in nature, while most studies only focus on the precision and adaptiveness of contents (Maier & Klotz, 2022).”

“Autonomy Experiment 1 used a 15-item SRL questionnaire developed by Dowson and McInerney (2004) as instrument for autonomy.”

“EPR borrows this logic by using instructional preferences and diagnostic data on learning conditions to recommend content or activities that are considered most beneficial for learners’ development (Zhang & Wang, 2022).”

“Furthermore, Schworm and Renkl (2006) found that in computer-supported example-based learning, providing instructional explanations could reduce students’ self-explanation activities, potentially impairing learning.”

“In the era when GAI significantly enhances the capabilities of EPR, it is especially important to examine its comprehensive effects on learners’ academic achievement and autonomy, given their rapid growth and increasing practical influence.”

“The social dimension examines students’ tendency to form interdependent learning relationships with those around them, including their inclination to communicate with, seek support from, and compare themselves to their teachers and classmates (Murase, 2015).”

“In professional contexts, Gruda (2024) demonstrated ChatGPT’s effectiveness in professional writing assistance.”

“For example, Ward et al. (2024) found that the use of AI tools improved academic performance among students by enhancing study habits, time management, and feedback mechanisms.”

“Some scholars have cautioned that extensive reliance on AI tools may foster technical dependency and undermine independent thinking and metacognitive monitoring (Hyde et al, 2024), yet rigorous, long-term investigations of these potential risks remain scarce.”

“In the field of STEM learning, students who develop SRL skills perform better on immediate tasks, such as problem-solving and computational thinking, and retain these skills over time, improving their ability to adapt and succeed in complex academic tasks (Tsai et al., 2013).”

“Extending the application to mobile learning environments, Drissi et al. (2024) developed GAMOLEAF, a gamified mobile learning framework for teaching Java data structures during COVID-19, and showed through a three-group comparison (N=90) that the integration of both gamification and personalized…”

“This capability facilitates in-depth exploration of help-related behaviors and the underlying learning mechanisms (Yang and Stefaniak, 2023).”

“Additionally, Dai et al. (2023) explored the role of adaptive testing technology in personalized recommendations for junior middle school students, revealing that recommending slightly more challenging content based on learner ability profiles significantly enhanced students’ scores and learning…”

“Building on the usage pattern categorization proposed by Du et al (2023) and Lai (2015), two primary AI-powered pedagogies in personalized learning were distinguished: (a) adaptive learning or personalized recommendation—wherein the system proactively creates learning pathways or resources based on…”

“Sancenon et al. (2022) conducted a randomized controlled trial with 43 Singaporean primary school students, implementing an adaptive recommendation system that customized worksheets based on learning analytics, and found that personalized content delivery significantly improved academic performance.”

“For SRL, Du et al. (2024) recruited 81 graduate students from an online course and employed a three-cycle design-based research approach to develop a recommender system for promoting SRL.”

“In terms of measurements, numerous tools have been developed for SRL, including questionnaires, interviews, think-aloud protocols, and online-based assessment systems (Wen et al., 2023).”

“[24] observed that mandated versus on-demand use of AI feedback by high school students had varied influences on their academic performance.”

“For instance, Dai et al. (2025) found that only students at the highest and lowest achievement levels had significant gains from the AI–assisted physics tutoring, while those in the middle level did not benefit equivalently.”

“Designs that put learners in control, such as tutor-like systems that hold back the final answer and ask for justification, or workflows with plan–revise cycles and reflective journals, tend to show stronger signals of self-regulation and ownership [92, 76, 83].”

“Ward et al. (2024) señalan que la retroalimentación generada por IAG puede mejorar los hábitos de estudio, y Dai et al. (2025) indican que, cuando se implementa adecuadamente, favorece la autorregulación y el aprendizaje adaptativo.”

“Notable research include simulating users in human-computer interaction (J. Wang & Fan, 2025; Xiang et al., 2024), and social simulations in social science (Park et al., 2023, 2024; X. Zhang et al., 2025).”

“While stronger empirical evidence would be required to establish causality, participant responses suggest that the interactive format supports ownership of the decision-making process [19] and reinforces the principle that responsibility for ethical judgement ultimately remains with the educator.”

“Such non-linear relation, where High level of integration did not directly interpret into highest gains, has been studied by Dai et al. [6].”

“However, their effectiveness could be compromised by challenges such as technical limitations, digital literacy deficits, and over-reliance on technology (Dai et al., 2025).”

“Nonetheless, Dai et al. (2025) warned that over-reliance on AI could reduce students learning independence.”

“…with it: given the same AI on demand, use split by prior achievement, the lowest achievers disengaging from it while the highest turned it to gain (Dai et al., 2025), though the same trial found that pairing compulsory use with hints helped the lowest achievers most, so the divergence tracks the…”

“However, the efficacy and practical value of these LLM-based solutions in operational, real-world classrooms, especially under the typical constraints of daily instruction, have not yet been sufficiently explored by empirical work(Dai et al. 2025).”

“AI contributes to these dimensions by offering personalized feedback, adaptive learning environments, dynamic assessments, and responsive interactions that are tailored to the individual needs of learners (Dai et al., 2025).”

“…randomized controlled experiments, and found that there were differences in learning autonomy and academic performance between students at different grade levels when they were “asked to use” and “used independently” GenAI, suggesting the criticality of the frequency and mode of AI use [ 7 ].”