Cognitive offloading and dependence on generative AI in learning

1. Vignette: a task completed and a learning that never arrived

The scene identifies no school, reports no grades, and cannot be read as an empirical finding. It condenses the pattern that withdrawal designs describe separately: the submission improves while the tool is open and grows poorer, or ceases to look like mastery, when the student is left alone. To treat the scene as data would invert the order of the article.

The question it opens is one of measurement, not of slogan. If the practice completed with the chatbot is graded, what is recorded is assisted performance. If the examination in which the tool is no longer present is graded, what is recorded is retained competence. That distinction —and not the volume of use— orders the evidence, the design of the help, and what an assessment may declare that someone has learned.

2. Problem: assisted performance versus retained competence

Assisted performance and retained competence are not two names for the same achievement. The first is what the student produces while the system takes part: practice with the tutor open, the essay written with the model, the problem that closes because the hint arrived in time. The second is what remains once that support is withdrawn: solving a kindred problem, revising a text of one’s own, justifying a step that nobody else drafted, detecting an error without the system pointing it out. Placing both under “learning outcomes” makes invisible the question that matters most for assessment.

The distinction matters because the grade is a proxy. If the assignment, and sometimes the examination, is done with the tool, the grade records the student-plus-system ensemble and overestimates what the person will do closed-book, or when the system fails. As a design inference, a ban does not follow. What follows is a refusal to treat assisted performance as if it were retained competence.

The contrast with quantitative synthesis is deliberate. As an empirical finding, Mo et al. (2026) estimate a large positive effect of ChatGPT on undergraduate outcomes (Hedges’ g = 1.14, SE = 0.185): of 5555 records, between 1 January 2023 and 31 May 2025 and across nine databases, they included 66 studies and 129 effect sizes. The critical question, which is not a result of Mo et al., is what share of those outcomes was measured with the tool present. The aggregate effect can be real and still remain silent about what is left when the tool is switched off.

The review question fixes population, concept, and context: students in secondary school and university; the offloading of cognitive processes onto generative artificial intelligence and its relation to competence without that help; tasks of learning and of assessment. The thesis that is anticipated is the thesis of the abstract. Unrestricted access can raise the assisted task and damage retained competence; design that conserves effort attenuates that cost; what decides is the mode of delegation, not the frequency; measuring only the assisted product looks at the wrong column. The evidence remains short, and very much supported by self-report.

3. Conceptual framework: cognitive offloading, productive effort, and agency

This section fixes terms; it is not a list of effects. Cognitive offloading is the use of an external resource to reduce the demand that would otherwise fall on the person’s own memory, reasoning, or evaluation. Gerlich (2025) treats it as a mediator between frequent use of artificial intelligence tools and critical thinking, in a general population and with a correlational design. What changes with generative models is not the idea of offloading, but the operation that is externalized: the system no longer stores a note the student made, but produces reasoning, synthesis, and, at times, the judgment.

Jose et al. (2025) enter only as a conceptual framework. The text that was recovered is not their abstract: no results are attributed to them, and the studies they cite are not imported. Read in that way, they formulate a paradox of enhancement and erosion. From cognitive load, a system can reduce extraneous load and, at the same time, thin the germane load that deep learning needs. From Bloom, automating recall and synthesis can coexist with a risk to judgment if the answer arrives already closed. From self-determination, a support that sustains agency is not the same as one that replaces it.

Dependence is used as Zhai et al. (2024) fix it in a systematic review of 14 articles: accepting the recommendations of a dialogue system without questioning them, when it is hard to judge reliability or how much trust to place. The result of that review, taken as a definition and not as a withdrawal experiment, is a preference for the quick solution and the heuristic shortcut. Dependence is not having opened the chatbot many times. It is having accepted its product without a verification that the student could have sustained alone.

Wang and Shan (2026) is a conceptual article, not a study with participants. They call the safety gap the divergence between assisted performance and the capacity to verify the result, above all when the system fails or is no longer there. Oracle models, they argue, remove the productive effort on which schemas depend, especially in the novice. They propose Socratic and adversarial architectures that do not hand over the solution: that is a design inference, not an effect. It names the gap in the vignette.

Agency joins the framework to the withdrawal test. Darvishi et al. (2024) ask whether the student learns from the assistance, so that the conduct continues when the prompt disappears, or leans on the assistance without learning. The detail belongs in section 5. Here the criterion is enough: agency is not a declared attitude, but justifying, revising, or solving without the scaffold. Four distinctions are not to be mixed again: offloading as less internal demand; dependence as uncritical acceptance; the safety gap as the distance between what is assisted and what is verifiable; agency as what a test without the tool observes.

4. Method: narrative review with an AI-free test criterion

The corpus was assembled on 12 October 2026 in the Crossref REST API: 16 queries, 60 rows, and filters for journal article and for publication from 2021. 960 records entered and 882 remained after deduplication. Automatic exclusion removed 39 without authors, 272 without a generative artificial intelligence term in the title, 118 without a cognitive or learning term, and 13 technical items off topic. Verification left 440 eligible. The seed had recorded 441; the difference of one record is variation of the index between runs, not a manual discard. The title filter required one generative-AI term and another term of cognition or of learning.

Manual screening gave priority to experiments, to tests after the support had been withdrawn, to meta-analyses, to reviews, and to scales with peer review. Left out were corrections, adoption surveys with no measure of learning or of offloading, and proposals without evaluation. From the seed came 20 works. A directed search, checked in Crossref, added Stadler et al. (2024), Darvishi et al. (2024), Jose et al. (2025), and Wang and Fan (2026). The corpus is 24: 21 with an abstract and 3 with title only —Hou et al. (2025), Liu et al. (2026), and Yan et al. (2026)—, and with no attributable results. Table 1 records the flow.

For each study it was recorded whether it measures (a) performance with artificial intelligence present, (b) performance after withdrawal or independent transfer, (c) self-report of dependence or of critical thinking, or (d) whether it is conceptual or a review. One article may occupy more than one cell. The narrative privileges (b), the only cell that separates the completed task from the learning that remains, and it reads (a) as assisted performance. Cell (c) informs the mode of delegation, with the reservation that belongs to self-report. Cell (d) lends names —paradox, safety gap, literacy in the interaction— and not effects.

This is not a systematic review and not a meta-analysis: there is no common effect. There was no independent double screening and no formal tool for risk of bias. The corpus is from Crossref alone. Wang and Fan (2026) were included on purpose, not as a learning outcome, but as a sample of how fragile early quantitative synthesis is. What follows argues about those abstracts; it does not claim to be an exhaustive map of the topic.

5. The withdrawal test: what happens when the tool is switched off

The decisive test is not how far the grade rises with the system switched on, but what remains when it is switched off. Table 2 gathers the contrasts that are available. They are not comparable with one another: a field study in upper-secondary mathematics, a quasi-experiment in writing, an inquiry of a single session, a note-taking study with thirty people, and an assistance to peer feedback that is not an open chatbot.

Bastani et al. (2025) separate the two columns in upper-secondary mathematics. With the tutor open, practice rose by 48% under GPT Base and by 127% under GPT Tutor. Once access was withdrawn, GPT Base stood 17% below those who had never had it; the guardrails mitigated that harm to a large extent. Without them, the model was used as a crutch. The same access can improve practice and worsen what comes afterwards, if during practice the tool did the cognitive work.

Chen (2026) measures the same dissociation in writing, with 180 undergraduates and 168 who completed eight weeks in six classes. The highest assisted mean was open collaboration (M = 4.02); the contrast with bounded support was imprecise (difference = 0.21, 95% confidence interval [0.06, 0.36], p = 0.064). In week 8, without AI, the bounded condition —a limit on delegation plus mandatory reflection— surpassed the open condition in quality (0.27), higher-order thinking (0.35), depth (0.42), and revision (0.39), with p from 0.050 to 0.063. Among 112 students who were exposed, the indirect path of offloading was estimated at −0.34 (95% confidence interval [−0.48, −0.21]). Self-regulated writing attenuated the link without erasing it. These are classroom associations, not definitive causal effects.

The rest is not one single effect. Darvishi et al. (2024), with 1625 students across ten courses, saw reliance on the feedback prompts instead of learning; self-regulation filled part of the gap, and the hybrid did not outperform AI alone. In Fan et al. (2025), ChatGPT improved the essay grade with no advantage in knowledge or in transfer. Stadler et al. (2024) do not withdraw the system: less load and less depth in the same session. Jiang et al. (2026) improve the outcome by delaying the support, without an examination taken with the chatbot closed. Yan et al. (2026) announce the comparison and contribute no effect.

The joint reading is an interpretation, not a common estimator. Where there is withdrawal, unrestricted access is associated with a better assisted product and with later competence that is worse or not improved, and limiting delegation attenuates that cost. Where there is no withdrawal, what appears is less depth with less load, or a better grade without better transfer. The pattern depends on the design and is not a single cause.

6. Perceived ease and metacognitive laziness

The student in the vignette did not say that she was ignorant of the topic. She said that she understood it on reading and could not produce it. Ease feels like learning. Lowering the load helps when what is removed is noise; it gets in the way when what is removed was the schema, and when fluency closes revision before its time.

Stadler et al. (2024) lower the load with a language model and, at the same time, argumentative quality relative to a search engine, without homogenizing the answers: the cost is less depth, not a single way of thinking. Fan et al. (2025) name metacognitive laziness. With no differences in intrinsic motivation, ChatGPT improved the essay grade and not knowledge or transfer, along distinct sequences of self-regulation. The label is theirs; the evidence of the abstract is that dissociation.

Preference widens the bias. In Chen et al. (2025) the worst posttest was the one that was preferred, because it was easy. Zhu et al. (2026) saw similar immediate benefits in both modes: maladaptive use is not noticeable at the moment. Brunnström and Palmqvist (2026) do not run an experiment. In their essay, the default output to “do machines think?” stayed in the multistructural phase of SOLO, and the qualitative phase required metacognitive interventions by the student. That is what they call literacy in the interaction with AI, and they warn, as a reflection, that unguided use could favor the person who already arrives with an advantage. Fluency is lived as comprehension. It is not.

7. Dependent and autonomous offloading

Checking a step that has already been attempted, and handing the task over so as not to attempt it, both add use and do not add the same learning. Zhu et al. (2026) separate dependent offloading —delegating the thinking— and autonomous offloading —a scaffold that does not let go of agency—. The outcomes are perceived. Across three waves, N = 589 among undergraduates and novice knowledge workers, the dependent mode was associated with a transfer of agency and with lower intrinsic motivation, and those paths with worse perceived outcomes of ability, creativity, processing, and judgment. The autonomous mode was associated with better motivation and better perceived outcomes. Metacognitive monitoring attenuated the link with the transfer of agency and did not attenuate the motivational cost. Both modes yielded similar immediate benefits. The question ceases to be whether one “uses AI.”

Jiang and Chen (2026) distinguish the order of what is externalized. Among 667 undergraduates in mainland China, the perceived value of AI was associated with lower-order offloading and with higher-order offloading through usage capability. With a high capacity to evaluate AI, that path ceased to be significant at the lower order and continued at the higher order: judging the output can restrain what is minor and not the judgment. Hou et al. (2025) only announce, in the title, a scale of dependence during problem solving. Without an abstract there are no scores.

There is no average user. Stojanov et al. (2024) obtained five profiles among 490 undergraduates and thirteen tasks: versatile low users, 38.2%; all-rounders, 10.4%; knowledge seekers, 16.5%, high in content, retrieval, and summary; proactive learners, 11.8%, high in feedback, planning, and self-evaluation; task delegators, 23.1%, reliant on the system to draft the submission. This is declared conduct, not learning. It is enough to keep from mixing, inside a single figure of use, the person who puts themselves to the test and the person who commissions the product that assessment grades.

Gerlich (2025) looks at a coarser grain, and not inside a classroom: 666 participants, a negative correlation between frequent use and critical thinking, mediated by offloading. The younger participants depended more and scored lower; a higher educational level was associated with better critical thinking apart from use. The design is correlational. Frequency is not irrelevant, and it is not the construct either: it sums delegations that are not the same. Zhai et al. (2024) describe the extreme as a preference for the shortcut. The variable that orders the risk is the stretch of thinking that is handed over, and whether the person who hands it over can still evaluate it.

8. Effort-preserving design: guardrails, delay, and scaffolding

Four designs come close to a test without the tool, and they are not interchangeable. The guardrails of Bastani et al. (2025) mitigated the harm to a large extent when access was withdrawn. The delay of Jiang et al. (2026) improved the outcome during use, without an examination taken with the chatbot closed. The limit with reflection in Chen (2026) held up better in week 8, with the reservations that belong to a classroom study. The intermediate assistance of Chen et al. (2025) surpassed maximal help on the posttest. The design inference is a bounded one: keeping one’s own production is associated with better competence than opening the tap.

A friendly wrapper is not enough. Melanou and Beege (2026), with 175 students and four hours, compared full scaffolding with a GCC template (goal, context, and constraints), light scaffolding, and a control. There were gains in knowledge, critical thinking, and reflective use in all of them, with no differences among the scaffolds; only a descriptive trend in higher-order knowledge. Joo et al. (2026), with 40 South Korean undergraduates working in Python, found process differences in favor of the metacognitive prompts, and a trend toward lower load, greater self-efficacy, and better performance. They mention a retention test and do not isolate a significant contrast for it. A trend is not retained competence.

In secondary school the counterevidence weighs more. Fütterer et al. (2026) randomly assigned 371 students in grades 7 to 9, across six sessions of 45 minutes in physics or English, to utility value, cognitive strategies, or standard ChatGPT. Perceived value evolved better in its own condition than in the strategy condition. None of them outperformed the control on effort, knowledge, or elaboration. An exploratory analysis linked a more substantive use with interest that was more sustained; there was no difference by subject. A support designed for self-regulation did not beat the ordinary chatbot. What accompanies better competence in the other studies is precise: not handing over the answer, delaying, and requiring reflection. Not just any script.

At early ages the withdrawal question is not closed. Liu et al. (2025) measured the usability of ChatGPT-MWPS in 52 fifth-grade students and across thirteen dimensions, favorable for reliability, information quality, and cognitive engagement: usability is not learning. Liu et al. (2026) and Yan et al. (2026) enter by title only and contribute no effects. As a design inference, not as a finding, in basic education, primary school included, replacing effort could be riskier, because verification is less fully formed. It is the novice logic of Wang and Shan (2026) carried over to age, and the corpus barely contains that evidence. The precaution is not to count as learned what was only done with help.

9. What the evidence does not yet support

There is no settled quantitative synthesis. Mo et al. (2026) estimate a large positive effect. The retraction note recorded as Wang and Fan (2026) withdraws a prior meta-analysis on ChatGPT, performance, perception of learning, and higher-order thinking, published online on 6 May 2025. The reason is not invented here: the text is the note, not a verdict. The lesson is caution toward early syntheses. A g of 1.14 compels one to take the assisted improvement seriously; a retraction compels one not to treat any meta-analysis of this generation as the last word on competence. Neither of the two answers the withdrawal question.

Self-report does not amount to a withdrawal either. Gerlich (2025), Zhu et al. (2026), Stojanov et al. (2024), and Jiang and Chen (2026) separate frequency, mode, profile, and order by means of what people declare. That immediate experience does not give maladaptive use away does not turn those surveys into Bastani et al. (2025), nor into week 8 of Chen (2026). Correlation is not causation. In Gerlich (2025), furthermore, educational level is associated with better critical thinking apart from use. The durations are short —four hours in Melanou and Beege (2026), eight weeks in Chen (2026), and six sessions in Fütterer et al. (2026)— and several designs fit inside one country or a few classes; Chen (2026) speaks of classroom associations.

“AI makes us stupid” does not come out of this corpus. There is later harm under a design without guardrails, a dissociation with p values at the edge, less depth with less load, and also large assisted effects, a secondary-school result in which scaffolding does not beat the ordinary chatbot, intensities of help that do not separate from one another, and a retracted meta-analysis alongside another that still stands. A conditional claim is admissible: unrestricted access can improve what is done with the tool and worsen what is done without it, and design moves the balance. A diagnosis of the age is not admissible.

10. Discussion: assessing what remains without the tool

Performance with the tool is bad news only if it is taken for competence. It can be high while the procedure is delegated, or while the tool is used as a shortcut without letting go of judgment. What separates the rows is the cell without the tool. The expectation is theoretical: it is inferred from Bastani et al. (2025), from Chen (2026), from the safety gap of Wang and Shan (2026), and from the perceived modes of Zhu et al. (2026), which do not fill that cell with a grade. Figure 1 is a working matrix, not a datum.

Four principles are a design inference, not a protocol that has been tried. Include moments without the tool. Assess process and verification, not only the product: Wang and Shan (2026) ask for this, and Chen (2026) separates it; Brunnström and Palmqvist (2026) ask, as a reflection, for the qualitative phase of SOLO. Teach students to direct and to evaluate the interaction, not only to formulate commands. Prefer tools that delay the answer or do not hand the solution over, without promising that any template will beat the ordinary chatbot: Melanou and Beege (2026) and Fütterer et al. (2026) rule that promise out. Neither a list, nor monitoring, nor knowing how to evaluate the output restores agency on its own, as Darvishi et al. (2024), Zhu et al. (2026), and Jiang and Chen (2026) show. The right-hand cell is observed by means of a task done without the system.

A further design inference, not a finding, concerns Mexico and Latin America. This corpus brings neither data nor regulation from those systems. The context that is assumed, without being measured, is one of large classrooms, unequal access, and tools calibrated above all in English. Unguided use can widen gaps: Brunnström and Palmqvist (2026) suggest, as a reflection, that the person who is already at an advantage draws more from a superficial chatbot. If the grade looks only at the assisted product, whoever has the system and knows how to direct it captures the grade. What follows is neither a program nor a ban, but a refusal to hand over the assisted column as if it were one’s own competence.

11. Limitations

The corpus is from Crossref alone: without Scopus, Web of Science, ERIC, or PsycInfo, absence here is not absence in the field. University samples predominate and, where the abstract states the country, Asian settings (mainland China, South Korea); no sample comes from Latin America. Three sources go by title only —Hou et al. (2025), Liu et al. (2026), Yan et al. (2026)— and Jose et al. (2025) only as a framework, because what was recovered is not their abstract. Also included are one study of a general population (Gerlich, 2025) and one human–computer interaction study with 30 participants (Chen et al., 2025). There was no double screening and no formal assessment of risk of bias. The tool changes faster than publication does.

There is almost no evidence from basic education, nor in Spanish. The vignette is composite and does not add a case to the corpus. Retained competence, moreover, is not the same scale across the twenty-four works: a grade without access, a rubric, a posttest, a transfer that does not differ, or an empty cell. Figure 1 lines them up inside one expectation. It does not prove that they are a single variable, and for that reason no common effect was estimated. The retraction recorded in Wang and Fan (2026) is part of that caution.

12. Conclusions

A task can turn out well and a learning can fail to arrive. In upper-secondary school, access without guardrails improved practice and worsened performance when the access was removed; a tutor that did not serve as a crutch mitigated the harm to a large extent. In writing, open collaboration looked stronger with the help in place, and the condition that limited delegation held up better without it, as a classroom association. In other designs there is less depth with less load, a better grade without better transfer, or a better posttest with intermediate help than with maximal help. The dissociation is a repeated one, and it is sensitive to design; it is not a single effect.

Frequency is a coarse signal, not the explanation. Delegating judgment, and leaning on the tool in order to go on judging, look alike in the immediate relief and separate afterwards. Profiles, scales, and a correlation mediated by offloading suggest as much, from self-reports that do not close the question of causality. The practical question is which stretch of thinking was handed over, and whether it reappears when the system is closed. Preserving effort, moreover, is not a matter of wrapping an oracle in a script: guardrails, delay, reflection, and incomplete assistance have some test behind them; the generic intensity of scaffolding, and usability in primary school, do not. For basic education the inference is one of precaution.

Assessment is where the distinction changes a grade. The assisted product is the column in which almost any mode can look good; what remains without the tool, and how it is verified, is the column that the mode does separate. This is a design inference, not a policy for Mexico or for Latin America. A meta-analysis of assisted outcomes read as learning, or a headline of universal decline, skips the withdrawal. More tests with the tool switched off are needed. That performance does not certify the competence that remains when the tool is switched off.

NEXTECH.IA Editorial Lab / Ingeniero Mitre

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), Article e2422633122. https://doi.org/10.1073/pnas.2422633122
2. Brunnström, L., & Palmqvist, L. (2026). AI-interaction literacy: reflections on how generative AI might be used to support self-regulated learning in higher education. Assessment & Evaluation in Higher Education. https://doi.org/10.1080/02602938.2026.2731261
3. Chen, X. (2026). Layer-sensitive cognitive offloading in generative AI-assisted writing: supported performance and independent no-AI outcomes. Frontiers in Psychology, 17, Article 1906199. https://doi.org/10.3389/fpsyg.2026.1906199
4. Chen, X., Ruan, K., Ju, K. P., Yap, N., & Wang, X. (2025). More AI Assistance Reduces Cognitive Engagement: Examining the AI Assistance Dilemma in AI-Supported Note-Taking. Proceedings of the ACM on Human-Computer Interaction, 9(7), 1–29. https://doi.org/10.1145/3757632
5. Darvishi, A., Khosravi, H., Sadiq, S., Gašević, D., & Siemens, G. (2024). Impact of AI assistance on student agency. Computers & Education, 210, 104967. https://doi.org/10.1016/j.compedu.2023.104967
6. 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
7. Fütterer, T., Bardach, L., Kuhn, J., Keller, S. D., & Gerjets, P. (2026). Enhancing School Students’ Self-Regulated Learning through Generative AI Support: A Randomized Controlled Trial. Educational Psychology Review, 38(1), Article 42. https://doi.org/10.1007/s10648-026-10133-8
8. Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006
9. Hou, C., Zhu, G., Sudarshan, V., Lim, F. S., & Ong, Y. S. (2025). Measuring undergraduate students' reliance on Generative AI during problem-solving: Scale development and validation. Computers & Education, 234, 105329. https://doi.org/10.1016/j.compedu.2025.105329
10. Jiang, Y., & Chen, H. (2026). AI-based cognitive offloading by students in higher education in China. Computers in Human Behavior: Artificial Humans, 8, 100327. https://doi.org/10.1016/j.chbah.2026.100327
11. Jiang, Y., Feng, X.-M., Liu, Y., Liu, B., & Bao, L. (2026). How does generative AI affect self-regulated learning in learners with varying knowledge levels? The role of chatbot design and support timing. Interactive Learning Environments. https://doi.org/10.1080/10494820.2026.2691907
12. Joo, S., Han, I., & Park, I. (2026). The effect of metacognitive prompts using generative AI on cognitive load, task performance, and self-efficacy in online self-regulated learning. Educational Psychology. https://doi.org/10.1080/01443410.2026.2618717
13. Jose, B., Cherian, J., Verghis, A. M., Varghise, S. M., S, M., & Joseph, S. (2025). The cognitive paradox of AI in education: between enhancement and erosion. Frontiers in Psychology, 16, Article 1550621. https://doi.org/10.3389/fpsyg.2025.1550621
14. Liu, J., Keane, T., Sun, D., Yang, Y., & Yang, Y. (2025). Design and evaluation of ChatGPT-MWPS: an AI-enhanced learning system for improving primary students’ mathematical word problem solving. Smart Learning Environments, 12(1), Article 68. https://doi.org/10.1186/s40561-025-00419-9
15. Liu, J., Zhang, Y., Li, W., Wang, Q., Niu, P., & Zhang, X. (2026). Adaptive vs. planned metacognitive scaffolding for computational thinking: Evidence from generative AI-supported programming in elementary education. Computers & Education, 241, 105473. https://doi.org/10.1016/j.compedu.2025.105473
16. Melanou, C., & Beege, M. (2026). Scaffolding Generative AI as a Tutor: A Quasi-Experimental Study of Learning Outcomes and Motivational, Cognitive and Metacognitive Processes. Education Sciences, 16(4), 651. https://doi.org/10.3390/educsci16040651
17. Mo, F., Huang, J., Yang, Y., Özen, Z., Maeda, Y., & Olenchak, F. R. (2026). Undergraduate students’ learning outcomes with ChatGPT: A meta-analytic study. Computers and Education: Artificial Intelligence, 10, 100536. https://doi.org/10.1016/j.caeai.2025.100536
18. Stadler, M., Bannert, M., & Sailer, M. (2024). Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry. Computers in Human Behavior, 160, 108386. https://doi.org/10.1016/j.chb.2024.108386
19. Stojanov, A., Liu, Q., & Koh, J. H. L. (2024). University students’ self-reported reliance on ChatGPT for learning: A latent profile analysis. Computers and Education: Artificial Intelligence, 6, 100243. https://doi.org/10.1016/j.caeai.2024.100243
20. Wang, H., & Shan, W. (2026). The safety gap: restoring productive struggle through pedagogically aligned generative AI. Frontiers in Education, 11, Article 1757622. https://doi.org/10.3389/feduc.2026.1757622
21. Wang, J., & Fan, W. (2026). Retraction Note: The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis. Humanities and Social Sciences Communications, 13(1), Article 528. https://doi.org/10.1057/s41599-026-07310-z
22. Yan, S., Yan, Y., Wang, S., Chen, S., Li, Y., & Ai, Z. (2026). Learning with AI, Not from AI: A quasi-experimental comparison of AI-assisted learning with versus without an interactive AI-generated error-correction phase on knowledge application and metacognition. International Journal of Educational Research, 139, 103047. https://doi.org/10.1016/j.ijer.2026.103047
23. Zhai, C., Wibowo, S., & Li, L. D. (2024). The effects of over-reliance on AI dialogue systems on students' cognitive abilities: a systematic review. Smart Learning Environments, 11(1), Article 28. https://doi.org/10.1186/s40561-024-00316-7
24. Zhu, Q., Li, X., Dong, Y., Chang, P., & Fan, M. (2026). Not all cognitive offloading is equal: distinguishing dependent and autonomous offloading to generative AI. Frontiers in Psychology, 17, Article 1878629. https://doi.org/10.3389/fpsyg.2026.1878629