Abstract
A growing body of research examines the cognitive effects of generative artificial intelligence, yet the population with the most at stake has been largely absent from this work. Kosmyna et al. demonstrated that adults using large language models for essay writing showed up to 55% reduced neural connectivity (dDTF magnitude) compared with unassisted writers, coining the term “cognitive debt.” Their findings are compelling, but the study is still a preprint, and their participants were adults, whose executive functions, unlike those of adolescents, were already largely built. Between ages 10 and 20, executive functions undergo rapid, experience-dependent maturation supported by the still-developing prefrontal cortex. This article discusses the Cognitive Offloading Paradox: generative AI is most useful for the cognitive tasks, such as planning, evaluating evidence, and constructing arguments, whose effortful practice is thought to drive executive function development during this critical window. The paradox identifies a specific, testable developmental risk that current research has overlooked. We advance it as a hypothesis. The distinction between AI as a scaffold (preserving cognitive effort) and AI as a substitute (replacing it) may determine whether the technology supports or undermines adolescent cognitive development, and should be the central organizing question for the next phase of research. The benefits these tools bring to learning are hard to dispute, and they are clearest when the AI scaffolds a student's thinking rather than doing the work for them.
1 Introduction
Kosmyna et al. (2025) fitted 54 participants with EEG caps and asked them to write essays under three conditions: using a large language model, using a search engine, or using only their own thinking. The LLM group showed the weakest neural connectivity of the three.
The study introduced “cognitive debt” to describe how delegating mental work to AI might spare effort now while accumulating costs later. However, every participant was between 18 and 39 years old, so their prefrontal cortices were close to or at maturity and their executive functions largely formed. The developing brain is another matter, and a recent fMRI study has begun to examine it directly: while using ChatGPT, children engaged the frontoparietal, salience, and attention networks that support cognitive control less than adults did (). Their creativity scores, however, did not differ from the adults'. The children recruited these networks less without any cost to what they produced—a difference in how the developing brain engages the networks, not in how well it performs. We return to this below.
Those most rapidly adopting generative AI for cognitive work are secondary school students and not adults in their twenties and thirties. Klarin et al. (2024) found that 52.6% of 17-year-olds in their sample reported using generative AI for schoolwork. These students are using AI for planning, writing, source evaluation, and argument construction at an age when those very capacities are still being built.
Whilst the cognitive debt that Kosmyna and colleagues described in adults may be temporary, in children and adolescents it could be developmental. That is the distinction this paper sets out to examine.
None of this has been demonstrated. What follows is a hypothesis. It draws on settled findings about executive-function maturation, about cognitive offloading, and about the experience-dependent development of the brain, and it makes predictions that can be subsequently tested. The evidence that would confirm or refute it does not exist yet, and a good part of this paper is about what that evidence would have to be.
2 Executive function maturation in adolescence
Executive functions (EFs) are the cognitive processes that allow people to plan, hold information in mind, resist impulses, and adapt when circumstances change. grouped them into three core components: working memory, cognitive flexibility, and inhibitory control, a structure supported by factor-analytic work on their unity and diversity (Miyake and Friedman, 2012). Together, these support the higher-order capacities that academic work demands, including reasoning, problem-solving, and self-regulated learning.
The developmental timeline of these functions is now well established. Tervo-Clemmens et al. (2023), in a study integrating four datasets with 10,766 participants aged 8 to 35, showed that executive functions follow a non-linear trajectory. Development is rapid between roughly ages 10 and 15, then gradually levels off, reaching adult-like performance around age 18 to 20. The pattern held across 23 different measures from 17 tasks, making it one of the most robust findings in developmental cognitive science. Nonetheless, the prefrontal systems behind these functions go on being refined into the mid-twenties. The steep part of the climb, though, belongs to adolescence.
This timeline tracks the biological maturation of the prefrontal cortex, which is among the last brain regions to complete structural development (; ). Synaptic pruning, myelination, and the strengthening of long-range neural connections in prefrontal regions continue through mid-adolescence and into the early 20's. Steinberg (2008) dual-systems model captures one consequence: the limbic system, which drives reward sensitivity and emotional reactivity, matures earlier than the prefrontal control system, creating an imbalance that defines much of adolescent behavior.
Maturation is experience-dependent and not just biological. argues that executive functions are trainable and shaped by experience, which implies that systematically reducing opportunities to practice them may have a negative impact on adolescent development. The principle is consistent with decades of neuroscience research showing that neural circuits consolidate through repeated activation (). The implication is that the cognitive demands placed on adolescents, including by their academic work, are essential steps in their developmental process.
If experience builds these functions, taking the experience away should weaken them, and the evidence bears that out. The deprivation side of early adversity, a relative lack of cognitive stimulation, predicts poorer executive function in childhood and adolescence (; Sheridan et al., 2017). When children were moved out of institutional care and into foster families in a randomized trial, the change altered the structural development of the prefrontal cortex and the tracts that connect it (Sheridan et al., 2022). This is severe deprivation, and the authors do not equate it with a teenager using ChatGPT. The prefrontal system is built by the demands made on it, so a sustained reduction in those demands is not developmentally neutral.
3 Why generative AI differs from previous cognitive tools
Cognitive offloading (i.e., using external tools to reduce internal processing demands) did not appear with the advent of generative AI. Risko and Gilbert (2016) provided the theoretical framework by positing that people offload when a task exceeds their perceived internal capacity and when an external tool is available. By this definition, a calendar is cognitive offloading. So is a calculator.
People worried about calculators too, and those concerns proved overblown. The difference is in what is being offloaded. A calculator takes over the arithmetic, but not the judgement of which calculation the problem needs; a text-to-speech tool takes over the decoding, but the reader still has to understand the words. A search engine takes over information retrieval but it is still up to the user to do the research and evaluate the sources given by it. An adolescent can use, or instruct, generative AI in such a way that it takes over planning, evaluation, composition, and the construction of an argument. Used in that way, what is offloaded are executive functions rather than peripheral operations, which is what distinguishes such use from earlier tools.
When a 16-year-old asks ChatGPT to “write an outline for my essay on climate policy,” the AI performs the task of holding multiple potential arguments in working memory, evaluating their relative strength, sequencing them into a logical structure, and making decisions about what to include and exclude (Kellogg, 2008). Every one of those operations exercises executive function. When the LLM does them instead, the engagement of those executive processes is likely reduced.
When a student accepts a finished output, what is left is the work of judging what the model has produced, whether the argument actually holds, whether the sources exist, whether the framing is the right one. That judgement is itself an executive and metacognitive act. Younger adolescents monitor their own thinking, and other people's, poorly, and they are inclined to trust text that reads fluently (Roebers, 2017). Accordingly, substitution works at both ends by removing the practice of the functions that build an argument, and leaning on the monitoring functions that are still the weakest.
Risko and Gilbert (2016) noted that offloading is more likely when internal cognitive demand is high and the individual perceives their own capacity as limited. Adolescents fit both criteria. Their executive functions are immature relative to the task demands of secondary education, and they are acutely aware of this gap. The structure of the situation pushes them toward substitutive rather than supplementary AI use.
4 The cognitive offloading paradox
We refer to this developmental form as the cognitive offloading paradox: generative AI provides the greatest immediate assistance with tasks that require executive function, and these same tasks, when performed effortfully, are thought to drive executive function development during adolescence. The technology is most useful for precisely the cognitive work that the developing brain most needs to do for itself. The framing sits on top of Risko and Gilbert's (2016) account of offloading; what makes it a paradox is that the person offloading has not yet built the capacities being offloaded. This developmental framing extends a broader discussion of a cognitive paradox of AI in education (Jose et al., 2025) and of cognitive offloading to AI more generally (Risko and Gilbert, 2016; ); our contribution is to specify its developmental form.
The paradox operates across several academic domains. In planning and task organization, AI can generate essay structures, project timelines, and revision schedules, displacing the working memory and sequencing demands that build planning capacity. In writing, AI can produce coherent paragraphs and arguments, reducing the need to coordinate multiple executive processes simultaneously. The Kosmyna et al. (2025) findings on reduced neural connectivity during LLM-assisted writing are directly relevant here, even though their participants were adults. Kellogg (2008) described skilled writing as one of the most cognitively demanding activities humans engage in, precisely because it requires the concurrent deployment of multiple EF components. In source evaluation, AI can summarize, compare, and assess the reliability of information, reducing the inhibitory control needed to resist plausible-sounding but weak claims and the flexibility needed to consider alternative interpretations.
5 The scaffold–substitute distinction
The distinction between scaffold and substitution draws on Vygotsky's (1978) concept of the zone of proximal development. A scaffold is an external support that enables a learner to accomplish something they cannot yet do alone, with the critical feature being that the learner still engages in effortful processing. A substitute removes the need for that processing entirely.
Current generative AI tools default to substitution. How far they substitute, though, depends on how they are used. A student may accept a finished product with little scrutiny, or may instead interrogate the model, weigh its answers, rework them, and check them against a teacher or peer before deciding (Wang, 2024). What is offloaded is then better described as a position on a continuum than as a fixed transfer of the work from student to model. The default settings nonetheless reward the low-effort path, and the pull toward it is strongest where workload and time pressure are high, which is when students report relying on these tools most (), an effect that appears among adolescents with weaker executive function as well (Klarin et al., 2024). This is, in part, a market-driven design, since users tend to prefer tools that minimize effort.
An AI tool functioning as a scaffold might ask a student questions about their essay rather than writing it. It might identify weaknesses in a draft and prompt the student to address them, rather than rewriting the passage. It might offer three possible structures for an argument and ask the student to evaluate which is strongest and why. In each case, the executive function demand is preserved or even heightened. The student still plans, still evaluates, and still decides.
coined the term “desirable difficulties” to describe learning conditions that make performance harder in the short term but produce better retention and transfer over time. Spacing, interleaving, and retrieval practice are all desirable difficulties. They work because they force the learner to do cognitive work. The concern with substitutive AI use is that it systematically eliminates desirable difficulties from academic tasks, optimizing for immediate performance at the expense of long-term learning and, in adolescents, at the potential expense of cognitive development itself.
The question to ask of any exchange is where the cognitive work happened. Did the student still generate the options, weigh them, and decide between them, or did the model hand over a finished answer to be approved? Did the exchange leave a process behind, a question asked, a draft criticized, a revision prompted, or only a finished product? They can be read from the record of an interaction, the prompts and the replies, and placed on a scale that runs from full scaffolding to full substitution, which is what studies will have to measure. Some tools have already been built in this direction, set up to ask questions, set goals, and prompt reflection instead of supplying answers (; ), and at least one controlled study with adolescents has set scaffolding use against substitutive use and against working alone (see Lindo and Cutad, 2024).
The effect, if there is one, will not fall evenly. Adolescents differ from one another, and several things are likely to shape how AI use bears on them: how old they are, since 13 is not 17; how strong their executive function and prior attainment already are; how digitally and AI-literate they are; what their home and school are like; and their socioeconomic context, which governs both the tools they can reach and the supervision that comes with them. This belongs to the paradox because the same tool, used the same way, may scaffold one student and substitute for another. It fits the finding that AI helps higher-order thinking more for students who already regulate their own learning well (Zhao et al., 2025). Accordingly, the students least able to manage their own learning, who are often the youngest, are the ones most likely to let the tool do the thinking, and the ones who, developmentally, can least afford to.
This places the paradox within a wider frame. Cognition is routinely distributed across tools and other people (; Sparrow et al., 2011), and adolescence is a period of heightened sensitivity to the social environment (). Extension into AI is unremarkable in a mature mind. The internal capacities that make such extension adaptive are still being built, and in developmental terms an external support has to be internalized before it can be relinquished (Vygotsky, 1978), so that leaning on a substitute during this window risks the internal substrate not forming. Whether AI scaffolds or substitutes is then settled within the socio-ecology of adolescence (), and not by the tool alone: the feedback a young person receives from the model, and from peers, teachers and parents, their prior digital experience, and the policies of their school. We sketch this ecology only in outline, because the evidence needed to map it does not yet exist; specifying and testing it is, we argue, the central task ahead.
The mode of use is important, and where the line falls between scaffold and substitute, for a given student, a given task, a given stage of development, is something we do not yet know.
6 The current evidence
Three terms often get used as if they were one, and they should be kept apart. “Cognitive debt” is Kosmyna et al.'s (2025) coinage, the idea that work handed to AI is paid for later. “Reduced neural engagement” means lower task-related activation while the tool is in use. “Reduced neural connectivity” means weaker coordinated activity between regions, measured by dDTF in that study and by within-network connectivity in the one we come to next. The argument needs only the modest version: a sustained drop in engagement, and in the practice it stands for, is important for development.
Two things limit how much weight the Kosmyna study can carry. It is a preprint, not yet through peer review, and a critical comment on its methods has already appeared (Stankovic et al., 2026). Furthermore, lower neural activity need not mean worse cognition; in studies of expertise it often marks more efficient processing, so that reading has to be carefully interpreted before alarms are raised.
scanned children and adults during a short creative exchange with ChatGPT. The children's control and attention networks, frontoparietal, salience, and dorsal attention, were less connected than the adults', which is what the paradox would predict. Their creativity scores, measured outside the scanner, were no different from the adults'. A clear difference in the brain came with no difference in what the children actually made. That is why reduced engagement cannot be read straight off as reduced ability, and it should hold back the strongest version of our argument as much as it motivates the question. Also, methodologically, the sample is small, the work is a preprint, and the participants were children of primary-school age, so the study speaks to the logic of the case and not to teenagers directly. The paradigm is also novel and not yet validated, there are no normative data for children of this age on the task, and the comparison is with adults; the study is best read as exploratory.
It would be one-sided to set out only the risks, because much of the evidence is encouraging. A meta-analysis of 69 experimental studies found that AI assistance tends to improve academic performance and higher-order thinking and to lower reported mental effort (). A second, of 29 experiments, found a moderate positive effect on higher-order thinking, stronger for problem-solving than for creativity, and stronger again for students who already self-regulate well (Zhao et al., 2025). A systematic review found that most studies report gains in self-regulated learning through personalized feedback and metacognitive support (Sardi et al., 2025), and individual studies show gains in adolescents and younger children when the tool is built to support regulation rather than replace it (Ng et al., 2024). A recent viewpoint in JAMA Pediatrics weighs the same promise and risk for adolescent health and cognition (Nagata et al., 2026).
The more careful of these studies draw the same line this paper does. The benefits gather where the AI prompts and questions while the student keeps doing the thinking; a controlled study that offloaded the lower-order work so students could concentrate on analysis and evaluation reported gains in critical thinking (). In contrast, students given a ChatBot improved their essays but gained nothing in knowledge or transfer, which the authors called “metacognitive laziness” (Fan et al., ), and the benefit to self-regulation held only where explicit metacognitive support was built in, and faded without it (Xu et al., 2025). These reported benefits sit on the other side of the same line of the proposed paradox. The practical task is to keep use on the scaffold side of it; the research task is to find where the line runs, for whom, and at what age.
Executive function and emotion regulation grow up together, on overlapping prefrontal and limbic circuitry (; Silvers, 2022), and weaker executive function tracks the emotion-regulation strategies adolescents fall back on (). If the kinds of AI use that bear on executive development are as common as they appear to be, they may bear on emotional development too. However, the direction of the link is disputed, and some longitudinal work runs the other way, with emotion regulation feeding later executive function (); and the AI use that matters here may be a different kind again, since teenagers increasingly talk to ChatBots for company and reassurance, which has prompted a developmental framework of its own (Liu and Yip, 2026). These are distinct questions, but not independent ones: because executive function and emotion regulation share developmental substrate, a mode of AI use that affects one may affect the other, and research on either would do well to measure both.
The evidence that bears directly on adolescents is, so far, correlational. Klarin et al. (2024) found associations between AI use and self-reported executive functioning, but a cross-sectional design cannot fix the direction of causation: weaker executive function may draw students to AI as readily as AI use might weaken it. reported a negative correlation between frequent AI use and critical thinking, with the youngest participants the most dependent and the lowest scoring, an age gradient that fits the proposed paradox. The Health Advisory on AI and Adolescent Wellbeing called adolescence a period of critical brain development that warrants particular attention, whilst acknowledging that longitudinal or experimental evidence on AI's cognitive effects at this age does not exist. The strands point the same way, but each is open to the same alternative reading, and none of them clearly addresses the question. The most prominent neuroscience study on AI and cognition excluded the population where the developmental stakes are highest.
The Cognitive Offloading Paradox is, at present, a theoretically motivated hypothesis. It is grounded in well-established findings on EF maturation, cognitive offloading, and experience-dependent neural development. It generates specific, testable predictions. But it remains unconfirmed.
However, our argument has clear limitations: the evidence is mostly cross-sectional and correlational. It rests on a small and fast-moving set of studies, two of the most relevant of them preprints. Nothing yet links a particular mode of AI use to long-term executive-function development. The published record probably leans toward negative findings, which draw more attention, so what is in print is not necessarily the whole picture. Furthermore, context such as teaching practice, parental guidance, a student's own AI literacy, will have an impact on findings. We do not think these limitations sink the argument; they set the terms on which it has to be tested.
7 A research agenda
Five lines of investigation would move this question from speculation to evidence.
First, the Kosmyna et al. (2025) paradigm should be replicated with adolescents aged 14 to 17, the period of most rapid EF maturation. If the neural connectivity reductions observed in adults are present or amplified in this age group, the paradox gains substantial empirical support. If they are not, that is equally important to know.
Second, longitudinal studies are needed that track executive function trajectories across an academic year in students with varying levels and modes of AI engagement. Standardized EF batteries administered at multiple time points, combined with detailed AI usage diaries, could test whether usage patterns predict differential developmental trajectories. Latent growth curve modeling would be the appropriate analytic framework.
Third, experimental work should directly manipulate the scaffold-substitute distinction. Randomly assign adolescent students to scaffold conditions (AI asks questions, provides feedback, prompts revision) vs. substitute conditions (AI produces outputs) vs. independent work, and measure cognitive engagement through behavioral, self-report, and, where feasible, neurophysiological indicators. The same studies should test how well students at different ages monitor and judge what the AI gives them, since that monitoring is itself a developing executive function; recording brain activity during evaluation, and not only during production, would show how the developing brain copes with output it did not build (Yan et al., 2025).
Fourth, effects are unlikely to be uniform across adolescence. Comparing early (10 to 13), mid (14 to 16), and late (17 to 20) adolescent groups would test whether the paradox's predictions are modulated by developmental stage, as Orben et al. (2022) work on sensitivity windows would suggest.
Fifth, adolescents themselves have something to contribute. RAND's American Youth Panel report (Schwartz and Diliberti, 2026) found 67% of students endorsed the statement that more AI for schoolwork harms critical thinking. This metacognitive awareness deserves systematic investigation through qualitative and participatory methods. These students understand something about their own cognitive experience that controlled experiments may miss.
8 Conclusion
The cognitive offloading paradox does not predict that generative AI will damage adolescent brains but instead it predicts that under certain conditions, specifically when AI substitutes for rather than scaffolds executive function processes, the technology may reduce the developmental stimulation that adolescent cognition requires during a sensitive period.
The appropriate response is not to ban AI from classrooms, which would be both impractical and counterproductive. Nor is it to wait for conclusive evidence before acting, since the technology is being adopted now and the developmental window does not pause for the publication cycle. The appropriate response is to investigate, with the methodological rigor and developmental specificity that the question demands, whether the paradox holds.
At minimum, the next wave of AI-cognition research must include the population that has been conspicuously absent from it so far. Studying what AI does to adult brains is interesting. Studying what it does to brains that are still being built is necessary.
Statements
Data availability statement
The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding authors.
Author contributions
JP: Writing – Conceptualization, Project administration, Supervision, Validation, Writing – original draft, Writing – review & editing. TK: Writing – original draft, Writing – review & editing.
Funding
The author(s) declared that financial support was not received for this work and/or its publication.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that Generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
1
AbbasM.JamF. A.KhanT. I. (2024). Is it harmful or helpful? examining the causes and consequences of Generative AI usage among university students. Int. J. Educ. Technol. High. Educ.21:10. doi: 10.1186/s41239-024-00444-7
2
AhmedS. P.Bittencourt-HewittA.SebastianC. L. (2015). Neurocognitive bases of emotion regulation development in adolescence. Dev. Cogn. Neurosci.15, 11–25. doi: 10.1016/j.dcn.2015.07.006
3
American Psychological Association (2025). Health Advisory on Artificial Intelligence and Adolescent Well-Being. Available online at: https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-ai-adolescent-well-being (Accessed April, 2026).
4
BjorkE. L.BjorkR. A. (2011). “Making things hard on yourself, but in a good way: creating desirable difficulties to enhance learning,” in Psychology and The Real World: Essays Illustrating Fundamental Contributions to Society, eds. M. A. Gernsbacher, R. W. Pew, L. M. Hough, and J. R. Pomerantz (New York, NY: Worth Publishers), 56–64.
5
BlakemoreS.-J. (2008). The social brain in adolescence. Nat. Rev. Neurosci.9, 267–277. doi: 10.1038/nrn2353
6
BlakemoreS.-J.ChoudhuryS. (2006). Development of the adolescent brain: implications for executive function and social cognition. J. Child Psychol. Psych.47, 296–312. doi: 10.1111/j.1469-7610.2006.01611.x
7
BlakemoreS.-J.MillsK. L. (2014). Is adolescence a sensitive period for sociocultural processing?Annu. Rev. Psychol.65, 187–207. doi: 10.1146/annurev-psych-010213-115202
8
BronfenbrennerU. (1979). The Ecology of Human Development: Experiments by Nature and Design. Cambridge, MA: Harvard University Press. doi: 10.4159/9780674028845
9
ChangD. H.LinM. P.-C.HajianS.WangQ. Q. (2023). Educational design principles of using ai chatbot that supports self-regulated learning in education: goal setting, feedback, and personalization. Sustainability15:12921. doi: 10.3390/su151712921
10
ChiuT. K. F. (2024). A classification tool to foster self-regulated learning with generative artificial intelligence by applying self-determination theory: a case of ChatGPT. Educ. Technol. Res. Dev.72, 2401–2416. doi: 10.1007/s11423-024-10366-w
11
ClarkA.ChalmersD. (1998). The extended mind. Analysis58, 7–19. doi: 10.1093/analys/58.1.7
12
DengR.JiangM.YuX.LuY.LiuS. (2025). Does ChatGPT enhance student learning? a systematic review and meta-analysis of experimental studies. Comput. Educ.227:105224. doi: 10.1016/j.compedu.2024.105224
13
DiamondA. (2013). Executive functions. Annu. Rev. Psychol.64, 135–168. doi: 10.1146/annurev-psych-113011-143750
14
FanY.„ Tang, L.„ Le, H.„ Shen, K.„ Tan, S.„ Zhao, Y. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. Brit. J. Educ. Technol.56, 489–530. doi: 10.1111/bjet.13544
15
FombouchetY.LannegrandL.LucenetJ. (2024). Relationships between emotion regulation strategies and executive functions in adolescence: exploring the effects of discrete emotions and age. J. Adolesc.96, 1239–1248. doi: 10.1002/jad.12334
16
GerlichM. (2025). AI tools in society: impacts on cognitive offloading and the future of critical thinking. Societies15:6. doi: 10.3390/soc15010006
17
GrinschglS.NeubauerA. C. (2022). Supporting cognition with modern technology: distributed cognition today and in an AI-enhanced future. Front. Artif. Intell.5:908261. doi: 10.3389/frai.2022.908261
18
HalseM.SteinsbekkS.BjørklundO.HammarÅ.WichstrømL. (2024). Emotions or cognitions first? longitudinal relations between executive functions and emotion regulation in childhood. Child Dev.95, 1508–1521. doi: 10.1111/cdev.14096
19
HongH.Vate-U-LanP.ViriyavejakulC. (2025). cognitive offload instruction with generative AI: a quasiexperimental study on critical thinking gains in english writing. Forum Ling. Stud.7, 325–334. doi: 10.30564/fls.v7i7.10072
20
Horowitz-KrausT.KenettY. N.LinkD.AshqarE.FarahR. (2025). Lower engagement of cognitive control, attention, modulation networks and lower creativity in children while using ChatGPT: An fMRI study. bioRxiv. doi: 10.1101/2025.11.07.687207
21
JohnsonD.PolicelliJ.LiM.et al. (2021). Associations of early-life threat and deprivation with executive functioning in childhood and adolescence: a systematic review and meta-analysis. JAMA Pediatr.175:e212511. doi: 10.1001/jamapediatrics.2021.2511
22
JoseB.CherianJ.VerghisA. M.VarghiseS. M.SM.JosephS. (2025). The cognitive paradox of AI in education: between enhancement and erosion. Front. Psychol.16:1550621. doi: 10.3389/fpsyg.2025.1550621
23
KelloggR. T. (2008). Training writing skills: a cognitive developmental perspective. J. Writ. Res.1, 1–26. doi: 10.17239/jowr-2008.01.01.1
24
KlarinM.HoffE.LarssonA.DaunkantaiteD. (2024). Adolescents' use and perceived usefulness of generative AI for schoolwork: exploring their relationships with executive functioning and academic achievement. Front. Artif. Intell.7:1415782. doi: 10.3389/frai.2024.1415782
25
KosmynaN.HauptmannA.YuanY. T.SituJ.LiaoX. H.BeresnitzkyA. V.et al. (2025). Your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv [Preprint]. arXiv: 2506.08872.
26
LindoM. R. S.CutadC. (2024). Impact of generative AI on self-regulated learning and cognitive offloading. IMCC J. Sci.4, 18–22. doi: 10.65931/c7k2p5z9
27
LiuC. H.YipT. (2026). Generative AI in adolescence: a developmental framework. JAMA Pediatr.180, 473–474. doi: 10.1001/jamapediatrics.2026.0032
28
MiyakeA.FriedmanN. P. (2012). The nature and organization of individual differences in executive functions: four general conclusions. Curr. Direct. Psycholog. Sci.21, 8–14. doi: 10.1177/0963721411429458
29
NagataJ. M.MemonZ.HuangO.MorenoM. A. (2026). Adolescent health and generative AI: Risks and benefits. JAMA Pediatr.180, 7–8. doi: 10.1001/jamapediatrics.2025.4502
30
NgD. T. K.TanC. W.LeungJ. K. L. (2024). Empowering student self-regulated learning and science education through ChatGPT: a pioneering pilot study. Brit. J. Educ. Technol.55, 1328–1353. doi: 10.1111/bjet.13454
31
OrbenA.PrzybylskiA. K.BlakemoreS.-J.KievitR. A. (2022). Windows of developmental sensitivity to social media. Nat. Commun.13:1649. doi: 10.1038/s41467-022-29296-3
32
RiskoE. F.GilbertS. J. (2016). Cognitive offloading. Trends Cogn. Sci.20, 676–688. doi: 10.1016/j.tics.2016.07.002
33
RoebersC. M. (2017). Executive function and metacognition: towards a unifying framework of cognitive self-regulation. Dev. Rev.45, 31–51. doi: 10.1016/j.dr.2017.04.001
34
SardiJ.DarmansyahC.andra, O.YulianaD. F.HabibullahYanto, D. T. P.ElizaF. (2025). How generative AI influences students' self-regulated learning and critical thinking skills? a systematic review. Int. J. Eng. Pedag.15, 94–108. doi: 10.3991/ijep.v15i1.53379
35
SchwartzH. L.DilibertiM. K. (2026). More Students Use AI for Homework, and More Believe it Harms Critical Thinking: Selected Findings from the American Youth Panel (RR-A4742-1). Santa Monica, CA: RAND Corporation. Available online at: https://www.rand.org/pubs/research_reports/RRA4742-1.html
36
SheridanM. A.MukerjiC. E.WadeM.HumphreysK. L.GarrisiK.GoelS.et al. (2022). Early deprivation alters structural brain development from middle childhood to adolescence. Sci. Adv.8:eabn4316. doi: 10.1126/sciadv.abn4316
37
SheridanM. A.PeverillM.FinnA. S.McLaughlinK. A. (2017). Dimensions of childhood adversity have distinct associations with neural systems underlying executive functioning. Dev. Psychopathol.29, 1777–1794. doi: 10.1017/S0954579417001390
38
SilversJ. A. (2022). Adolescence as a pivotal period for emotion regulation development. Curr. Opin. Psychol.44, 258–263. doi: 10.1016/j.copsyc.2021.09.023
39
SparrowB.LiuJ.WegnerD. M. (2011). Google effects on memory: cognitive consequences of having information at our fingertips. Science333, 776–778. doi: 10.1126/science.1207745
40
StankovicM.HircheE.KollatzschS.DoetschJ. N. (2026). Comment on: Your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing tasks. Preprint. Available online at: https://arxiv.org/abs/2601.00856
41
SteinbergL. (2008). A social neuroscience perspective on adolescent risk-taking. Dev. Rev.28, 78–106. doi: 10.1016/j.dr.2007.08.002
42
Tervo-ClemmensB.CalabroF. J.ParrA. C.FedorJ.ForanW.LunaB. (2023). A canonical trajectory of executive function maturation from adolescence to adulthood. Nat. Commun.14:6922. doi: 10.1038/s41467-023-42540-8
43
VygotskyL. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Cambridge, MA: Harvard University Press.
44
WangC. (2024). Exploring students' generative AI-assisted writing processes: Perceptions and experiences from native and nonnative English speakers. Technol. Knowl. Learn.30, 1825–1846. doi: 10.1007/s10758-024-09744-3
45
XuX.QiaoL.ChengN.LiuH.ZhaoW. (2025). Enhancing self-regulated learning and learning experience in generative AI environments: the critical role of metacognitive support. Brit. J. Educ. Technol.56, 1842–1863. doi: 10.1111/bjet.13599
46
YanL.Pammer-SchindlerV.MillsC.NguyenA.GaševićD. (2025). Beyond efficiency: empirical insights on generative AI's impact on cognition, metacognition and epistemic agency in learning. Brit. J. Educ. Technol.56, 1675–1685. doi: 10.1111/bjet.70000
47
ZhaoY.YueY.SunZ.JiangQ.LiG. (2025). Does generative artificial intelligence improve students' higher-order thinking? a meta-analysis based on 29 experiments and quasi-experiments. J. Intell.13:160. doi: 10.3390/jintelligence13120160
Summary
Keywords
adolescent development, cognitive offloading, Generative AI, developmental sensitivity, metacognitive monitoring
Citation
Pereira Campos J and Koff T (2026) Your brain on ChatGPT, but whose brain? The missing adolescent in AI-cognition research. Front. Dev. Psychol. 4:1885225. doi: 10.3389/fdpys.2026.1885225
Received
19 May 2026
Revised
14 June 2026
Accepted
15 June 2026
Published
08 July 2026
Volume
4 - 2026
Edited by
Stephen Butler, University of Prince Edward Island, Canada
Reviewed by
Tzipi Horowitz-Kraus, Technion Israel Institute of Technology, Israel
Vicent Mabirizi, Kabale University, Uganda
Updates
Copyright
© 2026 Pereira Campos and Koff.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Jorge Pereira Campos, jorge@drjorgecampos.com
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.