Different students, same chatbot: that gap is the whole story of learning to use generative AI well. It points toward a much older and better-studied skill than prompting β self-regulated learning, the ability to plan your approach to a task, monitor your own understanding as you go, and adjust when something isn’t working (Zimmerman, 1990; Panadero, 2017). Generative AI raises the stakes on a demand higher education already had before chatbots arrived: a longer shift toward self-directed learning, accelerated by years of online study, that had already outpaced how much self-regulation many students bring with them (Lodge et al., 2026). A framework commissioned by Australia’s higher education regulator, written by nineteen scholars including de Barba, gives the AI-era version of this skill its own name β hybrid metacognition, the regulation of thinking within a humanβAI system rather than inside a single head: evaluating what the model gives you, deciding when to rely on it and when to override it, catching the illusions of understanding it can produce before they harden into belief (Lodge et al., 2026).
The tool asks more of you
The intuitive story about AI is that it does the thinking so you don’t have to. The research on how people actually interact with these systems tells a more demanding story. Lev Tankelevitch and colleagues at Microsoft Research mapped the moments where working with generative AI leans hardest on metacognition β the ability to monitor and steer your own thinking β and found three of them stacked on top of each other in every session (Tankelevitch et al., 2023). Writing a prompt in the first place requires knowing your own goal clearly enough to put it into words and break it into steps, which is harder than it sounds; ask anyone who’s stared at a blank chat box unsure what they even want. Judging the output requires well-calibrated confidence in your own ability to tell a good answer from a fluent-sounding wrong one. And deciding whether to use the tool at all for a given task, over and over, requires the flexibility to keep updating that judgement as the task changes. This is the actual work of using the system, and it asks more of a novice than a blank page ever did.
What happens when that skill isn’t there yet
Self-regulation is a bundle of sub-skills, and students differ in which ones they’ve developed β some can set a clear goal but struggle to monitor their own progress toward it, others the reverse (Lodge et al., 2026). Two recent studies make the cost concrete, and one of them should give any teacher pause. Rania Abdelghani and colleagues sat 63 French middle-schoolers down with ChatGPT for a set of science tasks and measured, precisely, whether they could tell a good AI answer from a shallow one (Abdelghani et al., 2026). Their sensitivity to answer quality came out at essentially chance level β students accepted vague, low-value answers from the model about as readily as they accepted genuinely useful ones. Prior confidence in the subject didn’t protect them; if anything, students who trusted AI more evaluated its answers more carelessly. The one factor that reliably predicted better judgement was measured metacognitive regulation skill. The researchers called the resulting state, aptly, an illusion of understanding: the sense of having learned something, produced by a fluent answer rather than by actually closing the gap in your own knowledge.
The second study, with 480 medical students, found the same mechanism from a different angle. AI use was linked to both higher cognitive load and, through that load, weaker critical thinking, but the relationship wasn’t fixed. Students with stronger self-regulated learning skills were buffered against it; the tool cost them less because they were managing the interaction rather than being managed by it (Arshad et al., 2026). The pattern across both studies is the same: generative AI amplifies whatever capacity for self-direction a learner already brings to it, for better or worse.
What does help
Michael Gerlich ran a controlled experiment across 150 participants in Germany, Switzerland, and the UK, comparing unguided AI use against a structured five-step protocol, and the difference wasn’t subtle (Gerlich, 2025). Unguided use produced measurable cognitive offloading with no improvement in argument quality. The guided condition significantly reduced offloading, improved critical reasoning on an independently rated rubric, and narrowed the performance gap between more and less educated participants. Structured use helped the people who needed it most.
The protocol itself is simple enough to try this afternoon. Work out your own answer, or at least a rough direction, before you open the chat. Use the AI to gather information and check facts rather than to generate your argument for you. Rebuild your answer from what you found, in your own words, without copying. Then hand your draft back to the AI and ask it to attack it: what’s missing, what’s weak, what would a skeptic say. Revise once more yourself, and keep the authorship (Gerlich, 2025). Five steps, and the only one that involves the AI doing your thinking for you is the one that’s missing entirely.
Two smaller, more human-sized habits back this up. Paula de Barba’s team, building on the Monash pilot above, suggests making one small piece of your reasoning visible after any AI-assisted task: which suggestion you kept, which you threw out, and why (de Barba, 2026). It’s a tiny move that turns an invisible mental process into something you, or a teacher, can actually notice and build on. It’s also the same shift the TEQSA-commissioned framework urges at the institutional level: gathering evidence of learning at multiple points along the way, while the process is still visible (Lodge et al., 2026). And a Maastricht University survey of 199 students found that the ones who described AI as a “knowledgeable friend” or study partner, rather than an answer machine, consistently reported learning more from it than those who used it to fully outsource a task (Carroll, 2026). The frame you bring to the tool turns out to matter about as much as the tool itself.
Different skills, same chatbot
Neither of the Monash students was doing anything wrong, and neither needed a different chatbot. One was collecting answers. The other was running a deliberate learning process and had simply put the AI inside it, deciding what to ask, what to challenge, and what to carry forward. What separated them was several specific skills stacked together β judging your own goal clearly enough to prompt for it, telling a fluent answer from a right one, noticing when a sense of understanding hasn’t actually been earned (Tankelevitch et al., 2023) β skills students arrive with in different combinations and different amounts (Lodge et al., 2026). There’s an older, sharper way to put the same thing, drawn from decades of research on how regulation moves between a learner and whoever’s helping them: does the monitoring, evaluating, and deciding stay with the student, or does it move to whoever’s easiest to hand it to, and stay there? A more capable other, a tutor, a teacher, now a chatbot, can hold that regulation for a while as scaffolding, the way an adult ties a child’s shoelaces before the child can do it alone β but the scaffolding only builds the skill it’s standing in for if it eventually gets handed back (Hadwin & Oshige, 2011). One Monash student let the chatbot keep doing the noticing and judging. The other kept those jobs for themselves, using the tool for information while holding onto what that information meant. Gerlich’s structured protocol closed that gap furthest for the people who started behind, which is the encouraging part: handing regulation back is a practice, learnable the way any skill is learnable, and it can be taught (Gerlich, 2025).
Plan before you prompt, stay skeptical of fluency, make your reasoning visible, and keep the last word for yourself β older work, still teachable. A policy framework, a controlled experiment, a small pilot, and forty years of learning theory all keep landing on that same short list, which is what a field convergence looks like: learning to use generative AI well and learning to learn are turning out to be the same task, wearing a new interface.
References & Further Readings
Abdelghani, R., Murayama, K., Kidd, C., SauzΓ©on, H., & Oudeyer, P.-Y. (2026). The illusion of understanding: How middle-schoolers fail to regulate inquiry with ChatGPT in a science task (arXiv:2505.01106). arXiv. https://doi.org/10.48550/arXiv.2505.01106
Arshad, A., Lone, A., Arickswamy, L., Hassan, K., Alnaim, A. A., & AlFarhan, M. F. (2026). From AI use to critical thinking among medical students: A moderated mediation perspective on cognitive load and self-regulated learning. Frontiers in Psychology, 17. https://doi.org/10.3389/fpsyg.2026.1883053
Carroll, D. (2026, February 11). From shortcut to study partner: Student approaches to GenAI. EDLAB, Maastricht University. https://www.maastrichtuniversity.nl/news/edlab/eduminded/shortcut-study-partner-student-approaches-genai
de Barba, P. (2026, August 17). The missing piece of higher education’s AI response: Adaptive capabilities. Needed Now in Learning and Teaching. https://needednowlt.substack.com/p/the-missing-piece-of-higher-educations
Gerlich, M. (2025). From offloading to engagement: An experimental study on structured prompting and critical reasoning with generative AI. Data, 10(11), Article 172. https://doi.org/10.3390/data10110172
Hadwin, A., & Oshige, M. (2011). Self-regulation, coregulation, and socially shared regulation: Exploring perspectives of social in self-regulated learning theory. Teachers College Record, 113(2), 240β264. https://doi.org/10.1177/016146811111300204
Lodge, J. M., de Barba, P., Ainscough, L., Brazil, J. R., Broadbent, J., Ebbert, D., Frankland, S., Gabriel, F., GaΕ‘eviΔ, D., Hennicke, T., Lim, L.-A., Male, S. A., Mirriahi, N., Oliveira, E. A., Pacitti, H., RakoviΔ, M., Russell, J., Taylor-Griffiths, D., & Yang, S. (2026). Assuring quality learning in a gen AI-integrated future: The role of adaptive capabilities. Tertiary Education Quality and Standards Agency. https://www.teqsa.gov.au/guides-resources/resources/corporate-publications/assuring-quality-learning-gen-ai-integrated-future-role-adaptive-capabilities
Panadero, E. (2017). A review of self-regulated learning: Six models and four directions for research. Frontiers in Psychology, 8, Article 422. https://doi.org/10.3389/fpsyg.2017.00422
Tankelevitch, L., Kewenig, V., Simkute, A., Scott, A. E., Sarkar, A., Sellen, A., & Rintel, S. (2023). The metacognitive demands and opportunities of generative AI (arXiv:2312.10893). arXiv. https://doi.org/10.48550/arXiv.2312.10893
Zimmerman, B. J. (1990). Self-regulated learning and academic achievement: An overview. Educational Psychologist, 25(1), 3β17. https://doi.org/10.1207/s15326985ep2501_2