Co-funded by the European Union
AI as a Learning Partnera metacognitive approach for academic staff

Step 3 of 5 · The research frame

What does the research say?

You have now explored the question yourself and written a reflection. Here is the evidence base behind the method, and why it matters for how you design learning with AI.

94%of UK students

In the UK, 94% of students say they use generative AI to help with assessed work (Stephenson & Armstrong, 2026). With figures like that, what needs asking is how students use AI, and what it does to their learning.

The performance paradox

The most important finding from recent research concerns something subtler than cheating. AI can make students appear to learn more while they actually learn less.

The performance paradox

With AI, students' short-term performance on a task can improve: they produce better answers, faster. Their long-term learning often suffers at the same time, so performance and learning move in opposite directions (Yan et al., 2025; Lodge & Loble, 2026).

Key study: Bastani et al., PNAS 2025

In a randomised experiment with nearly a thousand high school math students in Turkey, researchers compared three groups: control (no AI), standard ChatGPT, and a pedagogically redesigned AI tutor that gave hints rather than answers.

With standard ChatGPT, students' grades on the practice problems went up by 48%. On a later exam without AI, they scored 17% lower than students who had never had access; the authors describe the AI being used as a “crutch”. With the hint-based tutor, practice grades rose too, and the harm to learning largely disappeared.

Why this happens: cognitive offloading

Polished, fluent answers from AI take away the effortful mental work that produces learning. This is cognitive offloading (Lodge & Loble, 2026). The task gets done, but the thinking that would have built lasting knowledge does not happen.

Helpful offloadingUsing AI to handle routine tasks, such as grammar checking, citation formatting or summarising background material, frees mental space for deeper thinking.
Harmful offloadingUsing AI to generate arguments, construct explanations or produce the work that should require the student to think. The effort that builds lasting understanding never happens.

The illusion of competence

In a study from Aalto University (Fernandes et al., 2026), AI made people better at the tasks and, at the same time, worse at judging their own performance. Participants who used AI scored higher on the problems but overestimated how well they had done by even more. The fluency of the output seems to make people feel they have understood more than they have.

The key insight for teachers

Students who use AI a lot are often the last to notice that they have not learned much. They feel confident and their submitted work looks good, so the gap only shows once AI is taken away: in an exam, in a discussion, or in any situation that calls for independent judgement.

Metacognition is the missing ingredient

What makes the difference, in study after study, is whether students stay mentally engaged while they use AI. Students who plan before they use it, monitor their understanding while they do, and reflect afterwards learn significantly more than those who simply hand the work over.

You did exactly this in Step 1. You wrote down your prior thinking, paused to ask whether AI was helping you think, and compared your position before and after. According to the research, those are the interventions that can keep learning intact even where AI is everywhere.

What well-designed AI tutoring looks like

In the Bastani study, the tutor that gave hints instead of answers largely preserved learning. In a Harvard RCT (Kestin et al., 2025), a pedagogically designed AI tutor even outperformed in-class active learning. Whether AI helps or harms learning depends on the pedagogical design around it.

Continue to Step 4, which turns the research into teaching: how to design the same metacognitive structure for your students.

References

  1. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
  2. Fernandes, D., Villa, S., Nicholls, S., Haavisto, O., Buschek, D., Schmidt, A., Kosch, T., Shen, C., & Welsch, R. (2026). AI makes you smarter but none the wiser: The disconnect between performance and metacognition. Computers in Human Behavior, 175, 108779. https://doi.org/10.1016/j.chb.2025.108779
  3. Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6
  4. Lodge, J. M., & Loble, L. (2026). Artificial intelligence, cognitive offloading and implications for education [Report]. University of Technology Sydney. https://doi.org/10.71741/4pyxmbnjaq.31302475 (PDF)
  5. Stephenson, R., & Armstrong, C. (2026). Student Generative Artificial Intelligence Survey 2026 (HEPI Report 199). Higher Education Policy Institute. https://www.hepi.ac.uk/reports/student-generative-ai-survey-2026/
  6. Yan, L., Greiff, S., Lodge, J. M., & Gašević, D. (2025). Distinguishing performance gains from learning when using generative AI. Nature Reviews Psychology, 4, 435–436. https://doi.org/10.1038/s44159-025-00467-5