
When using AI in learning, learners should focus on deepening their understanding of knowledge and actively engage in metacognitive processes such as evaluation, monitoring, and orientation, rather than blindly following ChatGPT’s feedback solely to complete tasks efficiently. (Fan et al., 2025, p. 2)
There is no shortage of articles warning about the dangers generative AI poses for student learning and the possibility of academic dishonesty. Distressing stories about likely student misuse of generative AI are becoming commonplace. It’s clear that both students and teachers can use AI in ways that are either beneficial or harmful to student learning. Teachers have to design course policies and learning activities in such a way that minimizes harmful AI use and leverages constructive use (if they choose to incorporate AI at all). Students must recognize when and how generative AI might help them learn and when it might hurt before following through with constructive AI use. In an ambitious laboratory study, Fan et al. (2025) documented one such AI pitfall, which they named metacognitive laziness. Then I got to experience it myself in one of my classes. First, let’s review the study.
Fan et al. (2025) randomly assigned college students to one of four groups: a control group, a ChatGPT group, a human group, and a checklist group. All students had two hours to complete a challenging writing assignment in which they had to read articles covering three topics in education and then synthesize the information into an essay on where they believed education would be in 10 years. They had to write the essay in English, which was a second language for all of them. The researchers provided students with a scoring rubric for the essay. Next, the students had one hour to revise their essays, and here is where the groups differed. The control[1] group revised without any additional resources. The ChatGPT group revised with access to ChatGPT 4.0. The human group had access to an expert writing teacher. Students in the checklist group received checklists that they could follow in their revisions. Researchers carefully recorded the steps students took in doing their revisions, such as how they used their resources and time spent revising or revisiting the readings. After students finished their revisions, the researchers gave them questionnaires to measure their knowledge of the readings, transfer of knowledge, and intrinsic motivation for the task. Finally, the researchers scored the essays according to the rubric.
The ChatGPT group scored significantly higher on their essays than all other groups, but none of the groups differed on the tests of knowledge, transfer, or intrinsic motivation. In other words, the ChatGPT group had written better essays but not gained any advantage in learning or motivation in doing so. What was telling was how the different groups used their resources. The ChatGPT group showed a tight loop between revising their essays according to the rubric and using ChatGPT. They were not monitoring their own understanding; they were monitoring how ChatGPT could help them meet the rubric. The human group showed a wider use of resources in their revisions than the ChatGPT group. They used feedback from their expert, revisited the readings, and engaged more in organizing and integrating the ideas from different sources during their revision. The other two groups also showed higher reliance on revisiting the readings and engaging in more elaborative processes.
Students in the ChatGPT group had offloaded the effortful task of monitoring and developing an understanding of the materials to ChatGPT, and they lost a sense of their own level of mastery in doing so. In other words, these students reduced the difficulty of the task but may have also developed a false sense of how well they learned and understood the material. It’s likely that many of them mistakenly believed that they learned more than they really did and never checked their actual level of knowledge. That’s metacognitive laziness. In the human group, the expert presumably pointed out to students not just the weaknesses of their essays, but suggested ways to address them by revisiting the sources and thinking about better elaboration and organization during the revision. The experts gave guidance about the process of revision that was missing in the ChatGPT group.
Here’s how I define metacognitive laziness. It occurs when learners offload cognitively demanding yet essential learning processes to generative AI, thereby undermining effective monitoring and self‑regulation of learning, and then never assess their actual knowledge. Offloading essential cognitive processes and not monitoring how it affects your learning is the “metacognitive” part. Failing to check your actual level of understanding is the “laziness” part. Cognitively demanding but essential learning processes include planning and goal setting, metacognitive monitoring and evaluation, selection and adaptation of learning strategies, and reflective integration and transfer of knowledge.
Now let me tell you about my experience with metacognitive laziness. I teach an upper-level cognitive psychology class, and the exams consist of prereleased identification items and essay questions. About two weeks before the exam date, I post a study guide with about 36 identification items and 10 essays. For the exam, I choose four identification items and two essay questions for students to answer. The exams are closed book. I warn the students that because they have so much time to prepare, the identification answers should include examples or research evidence in addition to the definition. The essays should be well organized and cite key research evidence. The students are free to discuss the terms and essays among themselves and ask me any questions. I create a discussion area on my LMS for questions. I’ve been using this exam format successfully for years
On my first exam of spring semester 2026, two of the identification terms I chose were stage assumption and abstract organism. Both terms are specific to cognitive psychology. Stage assumption is one of the key assumptions for the information-processing model of memory. Abstract organism is a key assumption of radical behaviorism. On the exam, about 15 percent of the class answered the stage assumption question by discussing Piaget’s stage theory of cognitive development, a widely influential and well-known theory. The problem is that I don’t teach Piaget at all, but if you put “stage assumption” into an AI model, you will likely get information about the best-known stage model in psychology. For abstract organism, about 15 percent gave me intelligent-sounding answers that had nothing to do with behaviorism—exactly what generative AI would give you if you entered the term as a prompt without any context. These students had clearly put the items into generative AI and used the answers without checking to see whether their answers made any sense in the context of the class. Many essay answers also showed evidence of metacognitive laziness. Students got zero points for these answers, and as a result, 24 percent of students failed the exam. Plus, they got a lecture from me about metacognitive laziness. The good news is that it was the first exam, and students had the rest of the semester to improve their grades, which virtually all of them did.
The quote that began this essay is a warning about metacognitive laziness. It undermines learning, leads to a false sense of confidence, and sets you up for failure. The problem is that it is a constant temptation. We have to teach students about it and other pitfalls generative AI creates lest they commit these errors in their professional lives. Students need to trust their teachers as their best source of information for exams. They have to realize that the context of the class matters, and general summaries from generative AI cannot provide that.
[1] These are my group labels, not the ones Fan et al. used.
Reference
Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
Stephen L. Chew is a cognitive psychologist who endeavors to translate cognitive research into forms that are useful for teachers and students. He is the recipient of multiple awards for his teaching and research. He was a long serving professor of psychology, now retired. Author contact: slchew@samford.edu.
The post Metacognitive Laziness and Student Use of Generative AI: A Classroom Example appeared first on The Teaching Professor.
