pith. sign in

arxiv: 2605.15850 · v2 · pith:GNR3B7VDnew · submitted 2026-05-15 · 💻 cs.CY · cs.AI· cs.HC

Access Timing as Scaffolding: A Reinforcement Learning Approach to GenAI in Education

Janne Rotter , Pau Benazet i Montobbio , Davinia Hern\'andez-Leo This is my paper

Pith reviewed 2026-06-30 19:31 UTC · model grok-4.3

classification 💻 cs.CY cs.AIcs.HC
keywords generative AIreinforcement learningeducational scaffoldingmetacognitionaccess timinghigher educationlearning outcomesproductive failure
0
0 comments X

The pith

A reinforcement learning agent that times GenAI access improves post-test performance and metacognitive accuracy over unrestricted or withheld access.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines whether controlling the timing of generative AI access can function as implicit scaffolding to support learning without inducing over-reliance. It implements this through a reinforcement learning agent whose reward function draws on metacognitive theory, cognitive load theory, and productive failure to decide when students may use GenAI. In a controlled lab study of 105 higher education students, the timed-access condition produced higher objective post-test scores and greater metacognitive accuracy than always-available access, while also cutting task errors and time on task relative to complete restriction. The gains occurred without any additional prompts or explicit scaffolding structures, indicating that access timing itself is a workable pedagogical lever.

Core claim

The authors show that an RL agent deciding when students may access GenAI, using a reward function grounded in metacognitive theory, cognitive load theory, and productive failure, yields better objective post-test performance and metacognitive accuracy than unrestricted access while reducing task errors and time on task compared with full withholding, in a mixed-methods study with 105 participants.

What carries the argument

A reinforcement learning agent that selects moments of GenAI access according to a reward function derived from metacognitive theory, cognitive load theory, and productive failure.

If this is right

  • Strategically timed GenAI access improves objective post-test performance compared with unrestricted use.
  • Timed access increases metacognitive accuracy relative to unrestricted access.
  • Timed access reduces task errors and time on task relative to complete withholding.
  • The timing approach outperforms both extremes without requiring explicit metacognitive prompts or structured scaffolding.
  • The method works with off-the-shelf GenAI tools and carries a low adoption barrier for educators.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Platform designers could embed similar agents to manage AI tool availability dynamically based on real-time student signals.
  • The timing-as-scaffolding principle may apply to other over-reliance risks such as search engines or code completion tools in targeted subjects.
  • Classroom trials over multiple sessions would reveal whether benefits hold outside a single controlled lab setting.
  • Different reward formulations could be tested to emphasize goals such as long-term retention or creative problem solving.

Load-bearing premise

The reward function based on those three theories produces access decisions that cause genuine improvements in learning rather than simply correlating with results in this lab task.

What would settle it

A replication in which students receive GenAI at the exact same times chosen by the RL agent but through random or fixed schedules instead of the theory-based rewards, and obtain equivalent gains in post-test performance and metacognitive accuracy.

Figures

Figures reproduced from arXiv: 2605.15850 by Davinia Hern\'andez-Leo, Janne Rotter, Pau Benazet i Montobbio.

Figure 1
Figure 1. Figure 1: Visualization of Experiment Design the students self evaluation and the pre-post test differ￾ence of the Metacognitive Awareness Inventory for Ar￾tificial Intelligence (MAI-AI) scales, a questionnaire to assess metacognitive awareness when working with AI, were treated as dependent variables. Additionally, time on each task, LLM logs and free text reflection on par￾ticipant’s work with the system were coll… view at source ↗
Figure 3
Figure 3. Figure 3: Boxplot comparison of the metacognitive ac [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Boxplot comparison of the objective post-test [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of GenAI usage patterns across [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of sentiment towards the own con [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of the final policy in the case of [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example screenshot of the ITS 22 [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
read the original abstract

In recent years, generative AI (GenAI) in educational settings has become ubiquitous in university students' daily lives, despite its potential to induce over-reliance, metacognitive disengagement, and diminished learning when used unrestrictedly. While most prior research has focused on how to pedagogically scaffold its usage, the question of when to allow off-the-shelf GenAI remains understudied and lacks pedagogically grounded empirical investigation. We treat access timing itself as a form of implicit scaffolding and operationalize it through a reinforcement learning (RL) agent that decides when students should access GenAI, with a reward function grounded in metacognitive theory, cognitive load theory, and productive failure. In a mixed-methods controlled lab study with N=105 higher education students, we compared the agent's effect on learning gains and metacognitive engagement to unrestricted and fully restricted use. Results show that strategically timed GenAI access under the reinforcement learning condition improved objective post-test performance and metacognitive accuracy compared with unrestricted access, while reducing task errors and time on task relative to complete withholding, thus outperforming both approaches without the need for explicit metacognitive prompts or structured scaffolding. However, no between-condition differences emerged on self-reported metacognitive awareness. Overall, timing of GenAI access therefore is a tractable, theoretically grounded, and scalable pedagogical strategy that improves over completely unrestricted and withheld access, compatible with off-the-shelf tools and potentially low adoption barrier. This opens up a new research area that explores how access timing can be facilitated by educators and implemented in human-AI learning system design.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript reports a mixed-methods controlled lab study (N=105 higher-education students) comparing three GenAI access conditions: unrestricted use, complete withholding, and access timing controlled by a reinforcement learning agent. The RL agent's reward function is described as grounded in metacognitive theory, cognitive load theory, and productive failure. The central empirical claim is that the RL condition produced higher objective post-test performance and metacognitive accuracy than unrestricted access, while reducing task errors and time on task relative to withholding, with no differences in self-reported metacognitive awareness. The authors conclude that access timing constitutes a tractable, theoretically grounded scaffolding strategy compatible with off-the-shelf tools.

Significance. If the results and the theoretical grounding of the reward function hold after full methodological disclosure, the work identifies a low-adoption-barrier lever (timing) for mitigating over-reliance on GenAI without requiring explicit prompts or structured scaffolds. The empirical three-condition design and the attempt to link RL decisions to established learning theories are strengths that could open a new line of inquiry on implicit scaffolding via access policies.

major comments (3)
  1. [Abstract / Methods] Abstract and Methods: The abstract asserts that the reward function is 'grounded in metacognitive theory, cognitive load theory, and productive failure' and that the RL condition 'outperformed both approaches,' yet supplies no formulation of the reward, no description of state/action spaces, no policy architecture, and no verification that the learned policy enacts the posited mechanisms (e.g., withholding after productive struggle). This information is load-bearing for the scaffolding interpretation; without it, the performance gains could arise from any beneficial timing heuristic rather than the claimed theoretical scaffolding.
  2. [Results] Results: No statistical tests, effect sizes, confidence intervals, or details on participant random assignment, task content, or pre-test equivalence are reported, despite the abstract claiming improvements on multiple objective measures. These omissions prevent assessment of whether the data support the between-condition claims.
  3. [Discussion] Discussion / Limitations: The manuscript does not include an ablation comparing the theory-derived reward against a non-theory-based timing rule (e.g., fixed delay or error-triggered access). Such a control is necessary to isolate whether the theoretical grounding, rather than any adaptive schedule, drives the observed advantages.
minor comments (2)
  1. [Abstract] The abstract states 'no between-condition differences emerged on self-reported metacognitive awareness' but does not specify the instrument or subscale used; this detail should be added for replicability.
  2. [Figures/Tables] Figure or table captions should explicitly state sample sizes per condition and whether error bars represent standard error or confidence intervals.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive and detailed feedback. We address each major comment below, indicating where we will revise the manuscript to improve clarity and rigor.

read point-by-point responses

  1. Referee: [Abstract / Methods] Abstract and Methods: The abstract asserts that the reward function is 'grounded in metacognitive theory, cognitive load theory, and productive failure' and that the RL condition 'outperformed both approaches,' yet supplies no formulation of the reward, no description of state/action spaces, no policy architecture, and no verification that the learned policy enacts the posited mechanisms (e.g., withholding after productive struggle). This information is load-bearing for the scaffolding interpretation; without it, the performance gains could arise from any beneficial timing heuristic rather than the claimed theoretical scaffolding.

    Authors: We agree that the specific formulation of the reward function, state and action spaces, policy architecture, and verification of alignment with theoretical mechanisms are essential to substantiate the scaffolding claims. The current manuscript provides only a high-level description. In the revised version, we will add a detailed subsection in Methods that includes the mathematical definition of the reward function (with explicit links to metacognitive theory, cognitive load theory, and productive failure), the state and action space definitions, the policy architecture, and any available analysis or examples showing how the learned policy implements the intended mechanisms such as withholding after productive struggle. revision: yes

  2. Referee: [Results] Results: No statistical tests, effect sizes, confidence intervals, or details on participant random assignment, task content, or pre-test equivalence are reported, despite the abstract claiming improvements on multiple objective measures. These omissions prevent assessment of whether the data support the between-condition claims.

    Authors: We acknowledge that the Results section as presented lacks the requested statistical details and supporting information on the experimental design. In the revision, we will expand the Results section to report all statistical tests (including appropriate omnibus and post-hoc tests), effect sizes, confidence intervals, the randomization procedure for participants, task content descriptions, and pre-test equivalence checks. This will enable full evaluation of the between-condition differences. revision: yes

  3. Referee: [Discussion] Discussion / Limitations: The manuscript does not include an ablation comparing the theory-derived reward against a non-theory-based timing rule (e.g., fixed delay or error-triggered access). Such a control is necessary to isolate whether the theoretical grounding, rather than any adaptive schedule, drives the observed advantages.

    Authors: We recognize the value of an ablation to isolate the contribution of the theory-derived reward. However, adding such conditions would require a new experimental design and additional participant recruitment, which exceeds the scope of a revision to the existing study. We will revise the Limitations and Future Work sections to explicitly note this gap and recommend it as a priority for subsequent research. The current three-condition design still provides comparative evidence for the RL approach relative to the two baselines. revision: partial

standing simulated objections not resolved
  • Request for an ablation study comparing the theory-derived reward against a non-theory-based timing rule, as this would require new data collection not feasible in the current revision.

Circularity Check

0 steps flagged

No significant circularity; empirical comparison of conditions with theory-grounded reward

full rationale

The paper describes an empirical mixed-methods lab study comparing three GenAI access conditions (RL-timed, unrestricted, fully restricted) with N=105 participants. The RL reward function is stated to be grounded in metacognitive theory, cognitive load theory, and productive failure, but the manuscript presents no equations, fitted parameters, or derivations. Results are reported as direct experimental outcomes on post-test performance, metacognitive accuracy, errors, and time on task. No self-citations are invoked as load-bearing uniqueness theorems, no ansatzes are smuggled, and no predictions reduce by construction to inputs. The derivation chain is therefore self-contained against external benchmarks (the controlled experiment itself).

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review; no equations, parameters, or entity definitions are supplied. The sole domain assumption visible is that the stated theories yield an effective reward function for access decisions.

axioms (1)
  • domain assumption Reward function grounded in metacognitive theory, cognitive load theory, and productive failure is appropriate for deciding GenAI access timing.
    Abstract states this as the basis for the RL agent but provides no further justification or validation.

pith-pipeline@v0.9.1-grok · 5826 in / 1242 out tokens · 41424 ms · 2026-06-30T19:31:51.793007+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. The Effortless Trap: Productive Struggle, AI, and the Illusion of Learning

    cs.CY 2026-06 unverdicted novelty 6.0

    Proposes a six-move framework (Prime, Probe, Point, Attach, Strengthen, Test) for learning with AI, using an 'effortless' diagnostic to avoid illusion of mastery, backed by cited evidence of design-dependent outcomes ...

Reference graph

Works this paper leans on

96 extracted references · 51 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    Collaborating with generative AI for learning?

    Davinia Hern´ andez-Leo et al. “Collaborating with generative AI for learning?” In:Proceedings of the 18th International Conference on Computer- Supported Collaborative Learning-CSCL. Interna- tional Society of the Learning Sciences. 2025, pp. 525–533.doi:10.22318/cscl2025.882858

    work page doi:10.22318/cscl2025.882858 2025

  2. [2]

    Experimental evidence on the productivity effects of generative artificial intelligence

    Shakked Noy and Whitney Zhang. “Experimental evidence on the productivity effects of generative artificial intelligence”. In:Science381.6654 (2023), pp. 187–192.doi:https://doi.org/10.1126/ science.adh2586

    2023

  3. [3]

    Promises and challenges of generative artificial intelligence for human learn- ing

    Lixiang Yan et al. “Promises and challenges of generative artificial intelligence for human learn- ing”. In:Nature Human Behaviour8.10 (2024), pp. 1839–1850.doi:10 . 1038 / s41562 - 024 - 02004-5

    2024

  4. [4]

    Digital Education Council.Global AI Student Sur- vey 2024. Tech. rep. Accessed: 2025-10-27. Digi- tal Education Council, 2024.url:https://www. digitaleducationcouncil.com/post/digital- education - council - global - ai - student - survey-2024

    2024

  5. [5]

    AI tools in society: Impacts on cognitive offloading and the future of critical think- ing

    Michael Gerlich. “AI tools in society: Impacts on cognitive offloading and the future of critical think- ing”. In:Societies15.1 (2025), p. 6.doi:10.3390/ soc15010006

    2025

  6. [6]

    Paris: OECD Publishing, 2026.doi:10

    OECD.OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Ed- ucation. Paris: OECD Publishing, 2026.doi:10. 1787/062a7394-en

    2026

  7. [7]

    Generative AI and the future of education: Ragnar¨ ok or reformation? A paradoxical perspective from management educa- tors

    Weng Marc Lim et al. “Generative AI and the future of education: Ragnar¨ ok or reformation? A paradoxical perspective from management educa- tors”. In:The international journal of management education21.2 (2023), p. 100790.doi:10.1016/ j.ijme.2023.100790

    work page arXiv 2023

  8. [8]

    Reinforcement learning in education: A literature review

    Bisni Fahad Mon et al. “Reinforcement learning in education: A literature review”. In:Informat- ics. Vol. 10. 3. MDPI. 2023, p. 74.doi:10.3390/ informatics10030074. 15

    2023

  9. [9]

    Plan More, Debug Less: Apply- ing Metacognitive Theory to AI-Assisted Program- ming Education

    Tung Phung et al. “Plan More, Debug Less: Apply- ing Metacognitive Theory to AI-Assisted Program- ming Education”. In:International Conference on Artificial Intelligence in Education. Springer. 2025, pp. 3–17

    2025

  10. [11]

    Improvement of AI- driven deep knowledge tracing algorithms

    Yan Li and Wai Yie Leong. “Improvement of AI- driven deep knowledge tracing algorithms”. In: 2024 International Conference on Intelligent Ed- ucation and Intelligent Research (IEIR). IEEE. 2024, pp. 1–6.doi:10 . 1109 / IEIR62538 . 2024 . 10959805

    2024

  11. [12]

    Inves- tigating the effects of LLM use on critical think- ing under time constraints: Access timing and time availability

    Jiayin Zhi, Harsh Kumar, and Mina Lee. “Inves- tigating the effects of LLM use on critical think- ing under time constraints: Access timing and time availability”. In:Proceedings of the 2026 CHI Con- ference on Human Factors in Computing Systems. 2026, pp. 1–21.doi:10.1145/3772318.3791796

    work page doi:10.1145/3772318.3791796 2026

  12. [13]

    Towards the pedagogical steering of large language models for tutoring: A case study with modeling productive failure

    Romain Puech et al. “Towards the pedagogical steering of large language models for tutoring: A case study with modeling productive failure”. In: arXiv preprint arXiv:2410.03781(2024)

    work page arXiv 2024

  13. [14]

    Generative AI in education: ChatGPT-4 in evalu- ating students’ written responses

    Jussi S Jauhiainen and Agustin Garagorry Guerra. “Generative AI in education: ChatGPT-4 in evalu- ating students’ written responses”. In:Innovations in Education and Teaching International62.4 (2025), pp. 1377–1394.doi:10.1080/14703297. 2024.2422337

    work page doi:10.1080/14703297 2025

  14. [15]

    A generative AI-based per- sonalized guidance tool for enhancing the feedback to MOOC learners

    ´Alvaro Becerra et al. “A generative AI-based per- sonalized guidance tool for enhancing the feedback to MOOC learners”. In:2024 IEEE Global Engi- neering Education Conference (EDUCON). IEEE. 2024, pp. 1–8.doi:10.1109/EDUCON60312.2024. 10578809

    work page doi:10.1109/educon60312.2024 2024

  15. [18]

    Effec- tiveness of crowd-sourcing on-demand assistance from teachers in online learning platforms

    Thanaporn Patikorn and Neil T Heffernan. “Effec- tiveness of crowd-sourcing on-demand assistance from teachers in online learning platforms”. In: Proceedings of the Seventh ACM Conference on Learning@ Scale. 2020, pp. 115–124.doi:10 . 1145/3386527.3405912

    work page arXiv 2020

  16. [19]

    Experimental evaluation of automatic hint generation for a logic tutor

    John Stamper et al. “Experimental evaluation of automatic hint generation for a logic tutor”. In: International Journal of Artificial Intelligence in Education22.1-2 (2013), pp. 3–17.doi:10.3233/ JAI-130029

    2013

  17. [20]

    Testing the multimedia principle in the real world: a compar- ison of video vs. text feedback in authentic mid- dle school math assignments

    Korinn Ostrow and Neil Heffernan. “Testing the multimedia principle in the real world: a compar- ison of video vs. text feedback in authentic mid- dle school math assignments”. In:Educational data mining 2014. 2014

    2014

  18. [21]

    The ef- fects of feedback interventions on performance: a historical review, a meta-analysis, and a prelimi- nary feedback intervention theory

    Avraham N Kluger and Angelo DeNisi. “The ef- fects of feedback interventions on performance: a historical review, a meta-analysis, and a prelimi- nary feedback intervention theory”. In:Psycholog- ical bulletin119.2 (1996), p. 254

    1996

  19. [22]

    Ed- ucation in the era of generative artificial intelli- gence (AI): Understanding the potential benefits of ChatGPT in promoting teaching and learning

    David Baidoo-Anu and Leticia Owusu Ansah. “Ed- ucation in the era of generative artificial intelli- gence (AI): Understanding the potential benefits of ChatGPT in promoting teaching and learning”. In:Journal of AI7.1 (2023), pp. 52–62.doi:10. 61969/jai.1337500

    2023

  20. [23]

    Vemprala, R

    Uday Mittal et al. “A comprehensive review on generative AI for education”. In:Ieee Access12 (2024), pp. 142733–142759.doi:10.1109/ACCESS. 2024.3468368

    work page doi:10.1109/access 2024

  21. [24]

    The promise and chal- lenges of generative AI in education

    Michail Giannakos et al. “The promise and chal- lenges of generative AI in education”. In:Be- haviour & Information Technology44.11 (2025), pp. 2518–2544.doi:10 . 1080 / 0144929X . 2024 . 2394886

    2025

  22. [25]

    The role of tutoring in problem solving

    David Wood, Jerome S Bruner, and Gail Ross. “The role of tutoring in problem solving”. In:Jour- nal of child psychology and psychiatry17.2 (1976), pp. 89–100

    1976

  23. [26]

    Application Research on Intelligent Teaching Feedback System Based on Reinforce- ment Learning Algorithm

    Cheng Feng. “Application Research on Intelligent Teaching Feedback System Based on Reinforce- ment Learning Algorithm”. In:Proceedings of the 2024 3rd International Conference on Artificial In- telligence and Education. 2024, pp. 563–566.doi: 10.1145/3722237.3722336

    work page doi:10.1145/3722237.3722336 2024

  24. [27]

    Metacognition and cognitive mon- itoring: A new area of cognitive–developmental inquiry

    John H Flavell. “Metacognition and cognitive mon- itoring: A new area of cognitive–developmental inquiry”. In:American psychologist34.10 (1979), p. 906.doi:10.1037/0003-066X.34.10.906. 16

    work page doi:10.1037/0003-066x.34.10.906 1979

  25. [28]

    A Study on the Metacognitive Awareness of Secondary School Students

    Sajna Jaleel and P Premachandran. “A Study on the Metacognitive Awareness of Secondary School Students”. In:Universal Journal of Educational Research4.1 (2016), pp. 165–172.doi:10.13189/ ujer.2016.040121

    work page arXiv 2016

  26. [29]

    An overview: Metacognition in education

    Mohsen Mahdavi. “An overview: Metacognition in education”. In:International Journal of Multidis- ciplinary and current research2.6 (2014), pp. 529– 535

    2014

  27. [31]

    A randomized controlled trial of interleaved mathematics practice

    Doug Rohrer et al. “A randomized controlled trial of interleaved mathematics practice.” In:Journal of Educational Psychology112.1 (2020), p. 40.doi: 10.1037/edu0000367

    work page doi:10.1037/edu0000367 2020

  28. [32]

    Measuring metacognitive knowledge, monitoring, and control in the pharmacy classroom and experiential settings

    Michelle L Rivers, John Dunlosky, and Adam M Persky. “Measuring metacognitive knowledge, monitoring, and control in the pharmacy classroom and experiential settings”. In:American journal of pharmaceutical education84.5 (2020), p. 7730.doi: 10.5688/ajpe7730

    work page doi:10.5688/ajpe7730 2020

  29. [33]

    Fostering metacognition to support student learning and performance

    Julie Dangremond Stanton, Amanda J Sebesta, and John Dunlosky. “Fostering metacognition to support student learning and performance”. In: CBE—Life Sciences Education20.2 (2021).doi: 10.1187/cbe.20-12-0289

    work page doi:10.1187/cbe.20-12-0289 2021

  30. [34]

    Rela- tion between intellectual and metacognitive skills: Age and task differences

    Marcel VJ Veenman and Marleen A Spaans. “Rela- tion between intellectual and metacognitive skills: Age and task differences”. In:Learning and indi- vidual differences15.2 (2005), pp. 159–176.doi: 10.1016/j.lindif.2004.12.001

    work page doi:10.1016/j.lindif.2004.12.001 2005

  31. [35]

    Beware of metacognitive lazi- ness: Effects of generative artificial intelligence on learning motivation, processes, and performance

    Yizhou Fan et al. “Beware of metacognitive lazi- ness: Effects of generative artificial intelligence on learning motivation, processes, and performance”. In:British Journal of Educational Technology56.2 (2025), pp. 489–530.doi:10.1111/bjet.13544

    work page doi:10.1111/bjet.13544 2025

  32. [36]

    The role of critical thinking on undergraduates’ reliance behaviours on generative AI in problem- solving

    Chenyu Hou, Gaoxia Zhu, and Vidya Sudarshan. “The role of critical thinking on undergraduates’ reliance behaviours on generative AI in problem- solving”. In:British Journal of Educational Tech- nology56.5 (2025), pp. 1919–1941.doi:10.1111/ bjet.13613

    2025

  33. [37]

    A comprehensive AI policy education framework for university teaching and learning

    Cecilia Ka Yuk Chan. “A comprehensive AI policy education framework for university teaching and learning”. In:International journal of educational technology in higher education20.1 (2023), p. 38. doi:10.1186/s41239-023-00408-3

    work page doi:10.1186/s41239-023-00408-3 2023

  34. [38]

    Analysing nontraditional stu- dents’ ChatGPT interaction, engagement, self- efficacy and performance: A mixed-methods ap- proach

    Mohan Yang et al. “Analysing nontraditional stu- dents’ ChatGPT interaction, engagement, self- efficacy and performance: A mixed-methods ap- proach”. In:British Journal of Educational Tech- nology65.5 (2025).doi:https://doi.org/10. 1111/bjet.13588

    2025

  35. [39]

    The effects of over-reliance on AI dialogue sys- tems on students’ cognitive abilities: a systematic review

    Chunpeng Zhai, Santoso Wibowo, and Lily D Li. “The effects of over-reliance on AI dialogue sys- tems on students’ cognitive abilities: a systematic review”. In:Smart Learning Environments11.1 (2024), p. 28.doi:10.1186/s40561-024-00316- 7

    work page doi:10.1186/s40561-024-00316- 2024

  36. [40]

    AI makes you smarter but none the wiser: The disconnect between per- formance and metacognition

    Daniela Fernandes et al. “AI makes you smarter but none the wiser: The disconnect between per- formance and metacognition”. In:Computers in Human Behavior175 (2025), p. 108779.doi:10. 1016/j.chb.2025.108779

    work page arXiv 2025

  37. [41]

    Enhancing self-regulated learning and learning experience in generative AI environments: The critical role of metacognitive support

    Xiaoqing Xu et al. “Enhancing self-regulated learning and learning experience in generative AI environments: The critical role of metacognitive support”. In:British Journal of Educational Tech- nology56.5 (2025).doi:10.1111/bjet.13599

    work page doi:10.1111/bjet.13599 2025

  38. [42]

    The metacognitive de- mands and opportunities of generative AI

    Lev Tankelevitch et al. “The metacognitive de- mands and opportunities of generative AI”. In: Proceedings of the 2024 CHI Conference on Hu- man Factors in Computing Systems. 2024, pp. 1– 24.doi:10.1145/3613904.3642902

    work page doi:10.1145/3613904.3642902 2024

  39. [43]

    Lessons learned and future directions of MetaTutor: Leveraging multichannel data to scaffold self-regulated learning with an in- telligent tutoring system

    Roger Azevedo et al. “Lessons learned and future directions of MetaTutor: Leveraging multichannel data to scaffold self-regulated learning with an in- telligent tutoring system”. In:Frontiers in Psy- chology13 (2022), p. 813632.doi:10 . 3389 / fpsyg.2022.813632

    work page arXiv 2022

  40. [44]

    Instructional Model Based on Intelligent Tutor Systems for the Development of 21st Century Competencies

    Claudia Cantero, Manuel Caro Pi˜ neres, and Juan C Giraldo Cardozo. “Instructional Model Based on Intelligent Tutor Systems for the Development of 21st Century Competencies”. In:Journal of Posthumanism5.7 (2025), pp. 2286–2302

    2025

  41. [45]

    Metacognitive over- load!: Positive and negative effects of metacogni- tive prompts in an intelligent tutoring system

    Kathryn S McCarthy et al. “Metacognitive over- load!: Positive and negative effects of metacogni- tive prompts in an intelligent tutoring system”. In:International Journal of Artificial Intelligence in Education28.3 (2018), pp. 420–438.doi:10 . 1007/s40593-018-0164-5

    2018

  42. [46]

    Leveraging deep rein- forcement learning for metacognitive interventions across intelligent tutoring systems

    Mark Abdelshiheed et al. “Leveraging deep rein- forcement learning for metacognitive interventions across intelligent tutoring systems”. In:Interna- tional Conference on Artificial Intelligence in Edu- cation. Springer. 2023, pp. 291–303.doi:10.1007/ 978-3-031-36272-9_24. 17

    2023

  43. [47]

    H., de Pater, I., Asay-Davis, X., Marcus, P

    Ido Roll et al. “Improving students’ help-seeking skills using metacognitive feedback in an intelli- gent tutoring system”. In:Learning and instruc- tion21.2 (2011), pp. 267–280.doi:10.1016/j. learninstruc.2010.07.004

    work page doi:10.1016/j 2011

  44. [48]

    Assessing metacognitive awareness

    Gregory Schraw and Rayne Sperling Dennison. “Assessing metacognitive awareness”. In:Con- temporary educational psychology19.4 (1994), pp. 460–475.doi:10.1006/ceps.1994.1033

    work page doi:10.1006/ceps.1994.1033 1994

  45. [49]

    Cognitive load during problem solv- ing: Effects on learning

    John Sweller. “Cognitive load during problem solv- ing: Effects on learning”. In:Cognitive science 12.2 (1988), pp. 257–285.doi:10 . 1016 / 0364 - 0213(88)90023-7

    1988

  46. [50]

    Cognitive of- floading

    Evan F Risko and Sam J Gilbert. “Cognitive of- floading”. In:Trends in cognitive sciences20.9 (2016), pp. 676–688.doi:10.1016/j.tics.2016. 07.002

    work page doi:10.1016/j.tics.2016 2016

  47. [51]

    Generative artificial in- telligence amplifies the role of critical thinking skills and reduces reliance on prior knowledge while promoting in-depth learning

    Guoqing Zhao et al. “Generative artificial in- telligence amplifies the role of critical thinking skills and reduces reliance on prior knowledge while promoting in-depth learning”. In:Educa- tion Sciences15.5 (2025), p. 554.doi:10.3390/ educsci15050554

    2025

  48. [52]

    Strategic offloading of delayed intentions into the external environment

    Sam J Gilbert. “Strategic offloading of delayed intentions into the external environment”. In: Quarterly journal of experimental psychology68.5 (2015), pp. 971–992.doi:10 . 1080 / 17470218 . 2014.972963

    work page arXiv 2015

  49. [53]

    The effects of acoustic turn-by- turn navigation on wayfinding

    Elliot P Fenech, Frank A Drews, and Jonathan Z Bakdash. “The effects of acoustic turn-by- turn navigation on wayfinding”. In:Proceedings of the human factors and ergonomics society annual meeting. Vol. 54. 23. SAGE Publications Sage CA: Los Angeles, CA. 2010, pp. 1926–1930.doi:10. 1177/154193121005402305

    2010

  50. [54]

    Navigational aids and spatial memory impair- ment: The role of divided attention

    Aaron L Gardony, Tad T Bruny´ e, and Holly A Tay- lor. “Navigational aids and spatial memory impair- ment: The role of divided attention”. In:Spatial Cognition & Computation15.4 (2015), pp. 246– 284.doi:10.1080/13875868.2015.1059432

    work page doi:10.1080/13875868.2015.1059432 2015

  51. [55]

    Google effects on memory: Cognitive conse- quences of having information at our fingertips

    Betsy Sparrow, Jenny Liu, and Daniel M Weg- ner. “Google effects on memory: Cognitive conse- quences of having information at our fingertips”. In:Science333.6043 (2011), pp. 776–778.doi:10. 1126/science.1207745

    2011

  52. [56]

    The “online brain

    Joseph Firth et al. “The “online brain”: how the Internet may be changing our cognition”. In:World psychiatry18.2 (2019), pp. 119–129.doi:https: //doi.org/10.1002/wps.20617

    work page doi:10.1002/wps.20617 2019

  53. [57]

    Human and AI collaboration in the higher education environment: opportunities and concerns

    Paul Atchley et al. “Human and AI collaboration in the higher education environment: opportunities and concerns”. In:Cognitive research: principles and implications9.1 (2024), p. 20.doi:10.1186/ s41235-024-00547-9

    2024

  54. [58]

    Trust and reliance on AI—An experi- mental study on the extent and costs of overre- liance on AI

    Artur Klingbeil, Cassandra Gr¨ utzner, and Philipp Schreck. “Trust and reliance on AI—An experi- mental study on the extent and costs of overre- liance on AI”. In:Computers in Human Behavior 160 (2024), p. 108352.doi:10.1016/j.chb.2024. 108352

    work page doi:10.1016/j.chb.2024 2024

  55. [59]

    Productive failure

    Manu Kapur. “Productive failure”. In:Cognition and instruction26.3 (2008), pp. 379–424.doi:10. 1080/07370000802212669

    2008

  56. [60]

    Productive failure in learning math

    Manu Kapur. “Productive failure in learning math”. In:Cognitive science38.5 (2014), pp. 1008– 1022.doi:https : / / doi . org / 10 . 1111 / cogs . 12107

    2014

  57. [61]

    Examining productive failure, pro- ductive success, unproductive failure, and unpro- ductive success in learning

    Manu Kapur. “Examining productive failure, pro- ductive success, unproductive failure, and unpro- ductive success in learning”. In:Educational Psy- chologist51.2 (2016), pp. 289–299.doi:10.1080/ 00461520.2016.1155457

    work page arXiv 2016

  58. [62]

    Learning from failure: A meta- analysis of the empirical studies

    Aubteen Darabi, Thomas Logan Arrington, and Erkan Sayilir. “Learning from failure: A meta- analysis of the empirical studies”. In:Educa- tional Technology Research and Development66.5 (2018), pp. 1101–1118.doi:10 . 1007 / s11423 - 018-9579-9

    2018

  59. [63]

    Ex- ploring mathematics problems prepares children to learn from instruction

    Marci S DeCaro and Bethany Rittle-Johnson. “Ex- ploring mathematics problems prepares children to learn from instruction”. In:Journal of experimen- tal child psychology113.4 (2012), pp. 552–568.doi: 10.1016/j.jecp.2012.06.009

    work page doi:10.1016/j.jecp.2012.06.009 2012

  60. [64]

    Practicing versus invent- ing with contrasting cases: The effects of telling first on learning and transfer

    Daniel L Schwartz et al. “Practicing versus invent- ing with contrasting cases: The effects of telling first on learning and transfer”. In:Journal of edu- cational psychology103.4 (2011), p. 759.doi:10. 1037/a0025140

    2011

  61. [65]

    Productive Failure and GenAI

    Michael Pak. “Productive Failure and GenAI”. In: Double Helix12 (2024).doi:10.37514/DBH- J. 2024.12.1.09

    work page doi:10.37514/dbh- 2024

  62. [66]

    Scientific Reports14(1), 23053 (2024)

    Yue Zhai and Behzad Nezakatgoo. “Evaluating AI-Powered Applications for Enhancing Under- graduate Students’ Metacognitive Strategies, Self- Determined Motivation, and Social Learning in English Language Education”. In:Scientific Re- ports15.1 (2025), p. 35199.doi:10.1038/s41598- 025-19118-z. 18

    work page doi:10.1038/s41598- 2025

  63. [67]

    Exploring artificial in- telligence in academic essay: higher education stu- dent’s perspective

    Agung Rinaldy Malik et al. “Exploring artificial in- telligence in academic essay: higher education stu- dent’s perspective”. In:International Journal of Educational Research Open5 (2023), p. 100296. doi:10.1016/j.ijedro.2023.100296

    work page doi:10.1016/j.ijedro.2023.100296 2023

  64. [68]

    Will generative AI replace teachers in higher educa- tion? A study of teacher and student perceptions

    Cecilia Ka Yuk Chan and Louisa HY Tsi. “Will generative AI replace teachers in higher educa- tion? A study of teacher and student perceptions”. In:Studies in Educational Evaluation83 (2024), p. 101395.doi:10 . 1016 / j . stueduc . 2024 . 101395

    2024

  65. [69]

    Embracing the future of Artificial Intelligence in the classroom: the relevance of AI literacy, prompt engineering, and critical thinking in modern education

    Yoshija Walter. “Embracing the future of Artificial Intelligence in the classroom: the relevance of AI literacy, prompt engineering, and critical thinking in modern education”. In:International Journal of Educational Technology in Higher Education21.1 (2024), p. 15.doi:10.1186/s41239-024-00448- 3

    work page doi:10.1186/s41239-024-00448- 2024

  66. [70]

    Reinforcement learning in education: A systematic literature review

    Anna Riedmann, Philipp Schaper, and Birgit Lu- grin. “Reinforcement learning in education: A systematic literature review”. In:International Journal of Artificial Intelligence in Education35 (2025), pp. 2669–2723.doi:10 . 1007 / s40593 - 025-00494-6

    2025

  67. [71]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto.Rein- forcement Learning: An Introduction. 2nd. Cam- bridge, MA: The MIT Press, 2018.isbn: 978-0- 262-03924-6

    2018

  68. [72]

    Axis: Generating explanations at scale with learnersourcing and machine learning

    Joseph Jay Williams et al. “Axis: Generating explanations at scale with learnersourcing and machine learning”. In:Proceedings of the third (2016) ACM conference on learning @ scale. 2016, pp. 379–388.doi:10.1145/2876034.2876042

    work page doi:10.1145/2876034.2876042 2016

  69. [73]

    Online optimization of teaching sequences with multi-armed bandits

    Benjamin Clement et al. “Online optimization of teaching sequences with multi-armed bandits”. In: 7th international conference on educational data mining. 2014

    2014

  70. [74]

    Towards prescriptive ana- lytics of self-regulated learning strategies: A rein- forcement learning approach

    Ikenna Osakwe et al. “Towards prescriptive ana- lytics of self-regulated learning strategies: A rein- forcement learning approach”. In:British Journal of Educational Technology55.4 (2024), pp. 1747– 1771.doi:10.1111/bjet.13429

    work page doi:10.1111/bjet.13429 2024

  71. [75]

    Reinforcement learning tutor better supported lower performers in a math task

    Sherry Ruan et al. “Reinforcement learning tutor better supported lower performers in a math task”. In:Machine Learning113.5 (2024), pp. 3023–3048. doi:doi.org/10.1007/s10994-023-06423-9

    work page doi:10.1007/s10994-023-06423-9 2024

  72. [76]

    A multimedia adaptive tutor- ing system for mathematics that addresses cogni- tion, metacognition and affect

    Ivon Arroyo et al. “A multimedia adaptive tutor- ing system for mathematics that addresses cogni- tion, metacognition and affect”. In:International Journal of Artificial Intelligence in Education24.4 (2014), pp. 387–426.doi:10.1007/s40593-014- 0023-y

    work page doi:10.1007/s40593-014- 2014

  73. [77]

    Koch et al.Cultural diversity in digital learning: Influences of cultural distance on perfor- mance and anxiety

    Nadine N. Koch et al.Cultural diversity in digital learning: Influences of cultural distance on perfor- mance and anxiety. Preprint, OSF Preprints. 2025. url:https : / / osf . io / preprints / psyarxiv / 6kjbm_v1

    2025

  74. [78]

    Fostering Metacognitive Awareness for Ef- fective Human–AI Collaboration in Learning

    Pau Benazet I Montobbio and Davinia Hern´ andez- Leo. “Fostering Metacognitive Awareness for Ef- fective Human–AI Collaboration in Learning”. In: Proceedings of the International Conference of the Learning Sciences (ICLS 2026). Irvine, USA: International Society of the Learning Sciences (ISLS), 2026

    2026

  75. [79]

    Bertoni, C

    E. Bertoni, C. Centeno, and R. Cachia.Social me- dia usage and adolescents’ mental health in the EU. JRC141047. Science for Policy Brief, Euro- pean Commission: European Commission, 2025

    2025

  76. [80]

    When knowing more means doing less: Algorithmic knowledge and digital (dis) engagement among young adults

    Myojung Chung. “When knowing more means doing less: Algorithmic knowledge and digital (dis) engagement among young adults”. In:Har- vard Kennedy School Misinformation Review6.5 (2025).doi:10.37016/mr-2020-186

    work page doi:10.37016/mr-2020-186 2025

  77. [81]

    [Online] Lesson plan

    Fiona Duffy et al.An Interactive Workshop on Body Image & Social Media for Schools and Youth Groups. [Online] Lesson plan. Bias project. 2022

    2022

  78. [82]

    [Online] Lesson plan

    Canada’s Center for Digital Media Literacy and TELUS.Body Image and Social Media: Escaping the Comparison Trap. [Online] Lesson plan. USE, Understand & Create: A Digital Literacy Frame- work for Canadian Schools. 2024

    2024

  79. [83]

    Proximal Policy Optimization Algorithms

    John Schulman et al. “Proximal policy op- timization algorithms”. In:arXiv preprint arXiv:1707.06347(2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  80. [84]

    Knowl- edge tracing: Modeling the acquisition of proce- dural knowledge

    Albert T Corbett and John R Anderson. “Knowl- edge tracing: Modeling the acquisition of proce- dural knowledge”. In:User modeling and user- adapted interaction4.4 (1994), pp. 253–278.doi: 10.1007/BF01099821

    work page doi:10.1007/bf01099821 1994

Showing first 80 references.