Beware of Metacognitive Laziness: Effects of Generative Artificial Intelligence on Learning Motivation, Processes, and Performance

Background: With the continuous development of technological and educational innovation, learners nowadays can obtain a variety of supports from agents such as teachers, peers, education technologies, and recently, generative artificial intelligence such as ChatGPT. In particular, there has been a surge of academic interest in human-AI collaboration and hybrid intelligence in learning.

Objectives: The concept of hybrid intelligence is still at a nascent stage, and how learners can benefit from a symbiotic relationship with various agents such as AI, human experts and intelligent learning systems is still unknown.

The emerging concept of hybrid intelligence also lacks deep insights and understanding of the mechanisms and consequences of hybrid human-AI learning based on strong empirical research.

Methods: In order to address this gap, we conducted a randomised experimental study and compared learners’ motivations, self-regulated learning processes and learning performances on a writing task among different groups who had support from different agents, i.e., ChatGPT (also referred to as the AI group), chat with a human expert, writing analytics tools, and no extra tool. A total of 117 university students were recruited, and their multi-channel learning, performance and motivation data were collected and analysed.

Results: The results revealed that: 1) learners who received different learning support showed no difference in post-task intrinsic motivation; 2) there were significant differences in the frequency and sequences of the self-regulated learning processes among groups; 3) ChatGPT group outperformed in the essay score improvement but their knowledge gain and transfer were not significantly different.

Conclusions: Our research found that in the absence of differences in motivation, learners with different supports still exhibited different self-regulated learning processes, ultimately leading to differentiated performance. What is particularly noteworthy is that AI technologies such as Chat- GPT may promote learners’ dependence on technology and potentially trigger metacognitive ”laziness”.

Introduction. 1 | INTRODUCTION In the 21st century, marked by constant technological evolution, artificial intelligence (AI) has become a catalyst for industrial and societal transformation. AI can automate many processes, significantly impacting the workforce and labour markets (Rane, 2023). It becomes essential for everyone to learn how to cooperate with AI and to capitalise on the new opportunities (Zarifhonarvar, 2023). Consequently, the importance of lifelong learning with human-AI collaboration is increasingly emphasised, signifying that everyone needs to continually acquire, adjust, and transfer knowledge and skills (Parisi et al., 2019), and more importantly, integrate the strengths of both humans and AI in the learning process (Järvelä et al., 2023). Following this vision, the concept of hybrid intelligence was proposed and several hybrid human-AI learning and regulation models were constructed (Holstein et al., 2020; Järvelä et al., 2023; Molenaar, 2022b). Akata et al. (2020) defined hybrid intelligence as a "combination of human and machine intelligence, augmenting human intellect and capabilities instead of replacing them and achieving goals that were unreachable by either humans or machines" (Akata et al., 2020, p. 19). Hybrid Intelligence is viewed as an evolving approach that addresses the limitations of data-driven AI (Järvelä et al., 2023), which often lacks interpretable and actionable insights, risk due to biased data and faces constraints in real-world applications (Ahmad et al., 2024). However, research into hybrid intelligence is still at a nascent stage (Molenaar, 2022b), and particularly, this field notably lacks deep insights and understanding of the mechanisms and outcomes of hybrid human-AI learning based on strong empirical research.

In the context of lifelong learning and hybrid intelligence, learners’ regulation plays a pivotal role, serving as a fundamental mechanism in an individual’s ability to engage effectively in learning (Taranto and Buchanan, 2020) and human-AI collaboration (Molenaar, 2022b). Self-regulated learning (SRL), as defined by Zimmerman (2000), involves self-generated thoughts, feelings, and behaviours directed toward achieving personal goals. The SRL model consists of three phases: forethought (where learners analyse tasks, set goals, and plan approaches, driven by motivational beliefs), performance (where they execute tasks, monitor progress, and apply self-control strategies to maintain focus and motivation), and self-reflection (where they assess performance, make attributions of success or failure, and adjust strategies for future tasks)(Zimmerman, 2000, 2002). Complementing SRL is metacognition, a term introduced by John Flavell in the 1970s, which refers to "thinking about thinking" or "cognition about cognition" (Flavell, 1979).

Metacognitive strategies in SRL, such as goal setting, self-monitoring, and self-evaluation, are essential for effective learning (Zimmerman, 2008). However, learners encounter diverse challenges during the process of regulation, stemming from factors such as inadequate metacognitive strategies, low achievement motivation, and task complexity (de Bruin et al., 2023; Russell et al., 2022; Wild and Grassinger, 2023). Hence, providing appropriate external support through different agents (i.e., teachers, peers, education technologies) is pivotal in facilitating the regulation process for learners. For instance, teacher feedback can aid learners in monitoring, evaluating, and validating their action plans (Brown and Palincsar, 2018). Peer support may assist learners in clarifying thoughts, rectifying misconceptions, and deepening comprehension (DiDonato, 2013). Furthermore, studies have emphasised the significance of intelligent and adaptive learning systems in bolstering SRL (Afzaal et al., 2021). Although most prior empirical studies do not directly pertain to AI, they establish a literary foundation that aids in understanding the various mechanisms and outcomes of SRL and inspires research interest in human-AI collaboration.

In recent years, leveraging AI technology to facilitate learning and regulation of learning has emerged as an important research area, and AI has shown both promises and issues when applied in education (Chiu et al., 2023; Grassini, 2023; Selwyn, 2022; Seo et al., 2021; Somasundaram et al., 2020). Particularly in late 2022, Generative AI (GenAI) technology such as ChatGPT garnered global attention for its ability to generate more coherent, systematic, and information-rich responses (Zhai, 2022). Due to ChatGPT’s capability to quickly and comprehensively assist learners in solving various complex problems and queries, writing essays, and learning programming, learning with GenAI has become an emerging research direction (Rahman and Watanobe, 2023).

Related work. Many studies posit that AI can enhance learning motivation. For instance, in a quasi-experimental study (Al- Abdullatif et al., 2023), learners who interacted with the task-oriented chatbot integrated with WhatsApp, showed higher motivation levels compared to a control group not using AI. Similarly, Lee et al. (2022) employed a quasiexperimental design to assess the effects of an AI-based chatbot used for after-class review, finding that learners in the AI group outperformed their counterparts in the control group in terms of academic performance, self-efficacy, learning attitude, and motivation. Yin et al. (2021) also found that learners in the AI chatbot-based learning environment attained higher levels of intrinsic motivation. Furthermore, a mixed-method study indicated that using ChatGPT positively influenced academic performance by enhancing learning motivation, suggesting effective use of AI-based tools can improve academic achievement by fostering motivation (Caratiquit and Caratiquit, 2023). However, these positive outcomes have not universally been echoed across all studies. For example, Fryer et al. (2017) utilised a 3 × 2 mixed design experiment to compare the effects of human and chatbot partners in oral language tasks. Contrary to the aforementioned studies, they found that interactions with human partners rather than chatbot were more effective in stimulating and sustaining task interest. A meta-analysis and systematic review of the effect of chatbot technology use in sustainable education also showed the use of chatbot technology could not significantly enhance learning motivation (Deng and Yu, 2023).

Previous research has provided valuable insights into the impact of AI on SRL processes. Several studies have focused on understanding learners’ perceptions, engagement, and SRL strategies when interacting with AI platforms.

For example, Clark et al. (2024) focused on non-science majors’ perceptions of a final exam facilitated by ChatGPT, underscored learners’ enhanced self-reflection and the importance of analysing AI-generated work, suggesting the potential role of AI in modulating SRL processes. Hwang et al.’s (2022) study on the smart chatbot application, Smart UEnglish, which analysed quantitative and semi-quantitative variables related to learners’ behaviours, revealed that AI significantly influenced learners’ behaviours in authentic English learning contexts, particularly during ’free talk’ and ’designed talk’ activities, underscoring the role of AI in facilitating complex conversation practice and enhancing learner engagement.

Method. 3 | METHODS 3.1 | Experimental Design and Settings Participants. A total of 117 university students (average age 22.61, SD=3.39, with 70% identifying as female and 55% undergraduates) participated in the experiment from July to September 2023. These participants came from a diverse array of disciplines. English was a second language for all the participants with the first language being [disclosed].

The participants were asked to complete a two-stage English reading and writing task as shown in Figure 1. The participants were randomly assigned to four experimental groups: one group did not have any support and finished the task by themselves (CN group, 30 participants); one group of learners were supported by ChatGPT 4.0 (AI group, 35 participants); one group of learners were supported by a human expert (HE group, 25 participants); and one group of learners had the support of the writing analytics toolkit named Checklist Tools (CL group, 27 participants).

Lab setting and research procedure. As shown in Figure 1, we conducted our experiment in the lab, where participants were required to complete the task following six steps: pre-task, stage 1 training, stage 1 reading and writing, stage 2 training, stage 2 revising, and post-task. In the lab, participants used a computer to complete pre-post-task questionnaires, watch training videos and complete the learning tasks. The first training video instructed the participants on how to use the learning tools in our learning environment. After watching the first training video, the participants began the 2-hour reading and writing task. The second training video introduced the participants to the support provided to them during the revision according to their experimental group assignment. After watching the second training video, the participants began the 1-hour revising task aiming to improve their essays. Once these tasks were completed, participants were asked to complete the post-test within one day.

Learning task. The participants, as English as second language speakers, were required to complete an English writing and revising task. In this task, we provided participants with reading materials on three topics: AI, differentiated teaching, and scaffolding teaching. The participants were expected to read these materials and write an essay that envisions the future of education in 2035 while integrating the three topics. Alongside these materials, a rubric was FIGUR E 2 Learning environment and learning tools provided for introducing the grading criteria of the essay. The participants could refer to this rubric while writing their essays, and researchers would score essays based on this rubric after the experiment.

3.2 | Four Learning Groups and Corresponding Learning Support Group Support Interaction Control group (CN group) - - ChatGPT group (AI group) ChatGPT 4.0, was restricted to task topics; using OpenAI API in our platform A chat frame; questions related to task could be asked, answers were limited to providing advice rather than directly generating an essay Human Expert group (HE group) Table 1 shows the different conditions of the four groups. the CN group maintained the same learning environment in both stage, without any additional support provided. the AI group received assistance from ChatGPT 4.0 (embedded in our platform user interfaces), which was trained and restricted to the content covered by our learning task (only conversations based on the learning task are allowed). The HE group received assistance from a proficient human expert who specialises in academic writing and academic writing education; participants could ask for help in polishing the content of the essay. The CL group had the support of writing analytics tools, which could provide feedback on (1) spelling and grammar, (2) academic style, (3) originality, and (4) rhetorical structure consistent with the genre the learners were asked to write in based on a GPT-based classifier of rhetorical categories designed in accordance with Author (2023) research, following Bloom’s taxonomy of cognitive domains (Bloom et al., 1956; Krathwohl, 2002).

Discussion. 5.2 | RQ2: Learners’ SRL process and Potential Metacognition Laziness Issue Our second research question (RQ2) focused on understanding the variations in SRL processes among learners interacting with different agents. Our results revealed significant differences during the revising stage, with the ChatGPT (AI), human expert (HE), and checklist (CL) groups engaging more extensively in elaboration, organization and orientation processes. This finding aligns with previous studies suggesting that AI and human tutors can facilitate interactive and personalized revision processes, moving beyond traditional reading-intensive approaches (Chen and Chang, 2024; Wei, 2023). Interestingly, the CL group also demonstrated a significant increase in evaluation processes, which can be attributed to the writing analytics diagnostic design of the checklist tools. Previous research has emphasized the importance of dynamic analysis of behaviour sequences to gain a deeper understanding of the SRL process (Li et al., 2023; Saint et al., 2022), and our findings contribute to this understanding by highlighting the impact of different agents aimed at supporting learning.

When comparing the AI group to other groups, frequency analysis and process mining revealed that the AI group exhibited SRL processes closely tied to the interactions with ChatGPT. Unlike the HE and CL groups, the AI group shows relatively fewer metacognitive processes (e.g., evaluation and orientation). Comparing process maps of the AI and HE groups, we found that the AI group’s SRL process centred on ChatGPT, while human teachers triggered more metacognitive process associations (e.g., transitions between orientation and evaluation). This highlights a crucial issue in human-AI interaction or hybrid intelligence: potential metacognitive laziness. In the context of human-AI interaction, we define metacognitive laziness as learners’ dependence on AI assistance, offloading metacognitive load, and less effectively associating responsible metacognitive processes with learning tasks.

Our definition of metacognitive laziness aligns with prior theoretical frameworks by Risko and Gilbert (2016) and Alter et al. (2007). Risko and Gilbert (2016) concept of cognitive offloading highlights how reliance on external resources can diminish learners’ engagement in metacognitive processes, leading to a habitual reduction in internal cognitive monitoring and self-regulation. Similarly, Alter et al. (2007) describe how experiencing cognitive disfluency encourages deeper analytical processing, indicating that bypassing such challenges—such as through dependence on AI-generated responses—can lead to a decrease in critical metacognitive activities. Together, these works reinforce our assertion that offloading metacognitive effort to AI tools results in less effective engagement with essential selfregulatory tasks, encapsulating the phenomenon we define as metacognitive laziness. Recent studies also highlight concerns related to this phenomenon. For example, Urban et al. (2024) found that using ChatGPT reduced learners’ perceived difficulty, leading to less effort in task solving. The same study also found that the perceived usefulness of ChatGPT was associated with learners’ self-evaluation judgements, resulting in higher inaccuracies in judgements of learning; therefore, Urban et al. (2024) argue that learners using AI tools should focus on effective metacognitive cues rather than relying on the perceived ease of AI-assisted problem-solving.

Conclusion. 7 | CONCLUSION In conclusion, our study highlights the potential of ChatGPT in improving essay scores, significantly outperforming other groups, including those guided by human experts. However, there were no significant differences in knowledge gain or transfer, indicating that while ChatGPT can enhance short-term task performance, it may not boost intrinsic motivation or long-term learning outcomes. The study also raises concerns about metacognitive laziness, where learners become overly reliant on AI, potentially hindering their ability to self-regulate and engage deeply in learning. This study contributes to the field of hybrid intelligence by revealing the potential and issues of learning with GenAI, and it calls for future research to deepen our understanding of how learners learn, regulate, collaborate, and evolve with AI.

Limitations. 6 | LIMITATIONS This study is subject to several limitations that may affect the generalizability and robustness of the findings. The observed lack of significant differences between groups could be attributed to constraints related to task duration and sample size. Therefore, further research with larger sample sizes and exploration of long-term effects on motivation and performance is necessary.