Most research on students and AI relies on self-reporting. Students fill out surveys, answer interview questions, tell researchers what they think they do when they open ChatGPT. The problem is obvious: what people say they do and what they actually do aren’t always the same thing. Corbin, Dawson, and Liu’s 2025 study on student-staff alignment around AI made a similar point, showing that agreement on paper doesn’t always translate to agreement in action.
Hawkins, Taylor-Griffiths, and Lodge (2025) tried something different. They gave 32 university students a timed essay task, told them to use generative AI, and recorded their screens the entire time. Then they brought each student back, played the recording, and asked them to narrate their own thinking. That kind of methodology is rare in this space, and it produced findings that move past the usual “students like AI” or “students cheat with AI” conversation. The strongest predictor of essay quality wasn’t AI fluency, confidence, or prior knowledge. It was feedback literacy.
How the Study Tracked AI Use in Real Time
The setup matters because it shapes how we should read the results. Hawkins et al. (2025) recruited 32 psychology students and gave them 25 minutes to write an essay on a topic they hadn’t seen beforehand. The students were explicitly asked to use generative AI during the task. Every screen interaction was recorded, and afterward, each student watched their own recording and explained what they were thinking at each moment. This video-stimulated recall method is borrowed from educational psychology research on teacher cognition. It bridges self-report and observation in a way most AI-in-education studies don’t even attempt.
From those interviews, Hawkins et al. (2025) identified four behavioral themes that followed a rough timeline during the task. Students typically started with “feed-forward” activities: asking AI for topic summaries, background information, or help understanding the essay prompt. They then shifted to “feedback” requests, asking AI to check their grammar, suggest synonyms, or paraphrase sentences. A smaller number moved into “feedback evaluation,” where they actively judged the quality of what the AI gave them. And half the participants, at some point during the task, deliberately stopped using AI altogether.
The Feed-Forward Trap
Nearly every student in the study opened by asking AI to generate ideas. Under the pressure of a 25-minute clock, that makes sense. Hawkins et al. (2025) found that this feed-forward phase dominated early interactions, with students treating AI as an idea generator and topic summarizer.
The trouble is that this phase rarely led to deeper engagement. Most feedback requests stayed at the surface level: fix my grammar, give me a better word, clean up this sentence. Fewer than 20% of students asked AI for substantive feedback on their argument or overall essay structure. A handful asked the chatbot to grade their work against the rubric, but that was the exception. The pattern Hawkins et al. (2025) describe looks a lot like what Fan et al. (2025) called metacognitive laziness: students defaulting to quick, low-effort interactions with AI because the tool makes it easy to skip the harder cognitive work.
How Students Evaluated AI Output
The most telling variation in the study was how differently students evaluated what AI gave them. Some compared AI-generated content against Google searches, journal articles, or a second chatbot. They treated AI output as a draft to be tested, not a product to be accepted. Others took what the chatbot produced and dropped it into their essay with minimal scrutiny.
Hawkins et al. (2025) note that negative evaluations of AI output, students calling the writing “repetitive,” “vague,” or “convoluted,” often pushed those students to revise the content themselves or abandon AI for that section entirely. That’s a productive failure. The students who noticed the AI wasn’t good enough were the same students who ended up writing stronger essays. This connects directly to Shaw and Nave’s 2026 work on cognitive surrender: the students most at risk aren’t the ones who use AI. They’re the ones who stop thinking once AI starts producing.
Why Half the Students Stopped Using AI
Half the participants actively stepped away from AI at some point during the task. Their reasons varied. Some cited academic integrity concerns. Others wanted the essay to sound like them. Several worried that reading too much AI-generated text would cause them to unconsciously mimic it, a concern worth taking seriously. Students are developing an intuitive sense of how AI can reshape their voice, even if they can’t articulate it in formal terms.
Hawkins et al. (2025) frame the student-AI dynamic through a co-regulation lens, where AI influences cognition, motivation, and behavior throughout the task. But they’re candid about the risks. As they put it, “without SRL skills grounded in self-efficacy and a motivation to learn, AI operates more like a student than a student tool” (p. 11). This suggests that for students who lack self-regulated learning skills, AI doesn’t assist the writing process. It replaces parts of it.
The researchers also raise a concern about unsupervised contexts, arguing that most AI-in-education research “ignores the unsupervised co-regulated learning environment students enter outside of the classroom when they interact with AI chatbots like ChatGPT that aren’t specifically designed to offer feedback in line with a pedagogical framework” (p. 11). Gerlich’s 2025 study on cognitive offloading and critical thinking found similar dynamics: the less structured support students had, the less their independent thinking held up.
Feedback Literacy as the Strongest Predictor
Carless and Boud (2018) define feedback literacy as “the understandings, capacities, and dispositions needed to make sense of information and use it to enhance work or learning strategies” (p. 1315, cited in Hawkins 2025 p. 2). In this way, feedback literacy, as Hawkins et al argue, “encompasses the skills to seek out, evaluate, and apply feedback to a process or task” (p. 2).
The quantitative side of the study showed that feedback literacy was the strongest predictor of essay performance. It outperformed prior knowledge, task confidence, effort, and even how often students used AI. Students who scored higher on feedback literacy didn’t necessarily use AI less. They used it differently: asked better questions, evaluated responses critically, and knew when to stop relying on the tool.
This finding aligns with what Hawkins et al. (2025) recommend: that institutions focus on building feedback literacy skills “that can be applied across domains so that students are prepared to effectively interact with, evaluate, and co-regulate their learning and productivity, no matter the context or technology” (p. 12). I’d agree with the recommendation, but I’d add a qualifier. Feedback literacy isn’t something you build in a single workshop or a one-off module. It develops through repeated exposure to meaningful feedback cycles, and most students don’t get enough of those even without AI in the picture.
Limitations and What Comes Next
Hawkins et al. (2025) are upfront about the constraints. Thirty-two students from a single psychology program, a 25-minute task with no real grade consequences, and an artificial setting where students were told to use AI. The findings are suggestive, not definitive. But the methodology is strong enough to justify replication, and the core insight, that feedback literacy shapes the quality of AI use, deserves to be tested across disciplines and at scale.
For educators, the practical lesson is clear enough. If we’re going to invest time in AI-related instruction, evaluation skills should come first. A student who can recognize when AI output is shallow, repetitive, or off-target will always outperform one who knows all the right prompts but can’t judge what comes back. The skill that matters most here isn’t technical, it’s critical.
References
- Carless, D., and D. Boud. 2018. “The Development of Student Feedback Literacy: Enabling uptake of Feedback.” Assessment & Evaluation in Higher Education 43 (8): 1315–1325. doi:10.1080/02602938.2018.1463354.
- Corbin, T., Dawson, P., Nicola-Richmond, K., & Partridge, H. (2025). ‘Where’s the line? It’s an absurd line’: Towards a framework for acceptable uses of AI in assessment. Assessment & Evaluation in Higher Education, 50(5), 705-717. https://doi.org/10.1080/02602938.2025.2456207
- 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.
- Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), Article 6. https://doi.org/10.3390/soc15010006
- Hawkins, B., Taylor-Griffiths, D., & Lodge, J. M. (2025). Summarise, elaborate, try again: Exploring the effect of feedback literacy on AI-enhanced essay writing. Assessment & Evaluation in Higher Education. https://doi.org/10.1080/02602938.2025.2492070
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646