Lower engagement of cognitive control in children on ChatGPT

Abstract

Generative artificial intelligence (AI) can scaffold idea generation, but how it engages control, attention, and memory systems in the developing versus mature brain, and how this relates to creativity, remains unclear. We compared 6–7-yr-old children and adults during a naturalistic 5-min co-creative dialog with ChatGPT using the same prompt, in-scanner conversation, undergoing functional magnetic resonance imaging scanning and their dialogs were recorded. Functional connectivity was computed within and between seven large-scale networks: cingulo-opercular, default mode, memory retrieval, frontoparietal (FP), salience, ventral attention, and dorsal attention (DAN). Creativity scores, assessed outside of the scanner, did not differ between groups. Adults showed stronger within-network connectivity in control/attention systems, most robustly in FP, with additional effects in salience and DAN. Adults also exhibited greater cross-network integration, particularly FP-salience and FP-DAN couplings, alongside broader adult-greater effects among memory- and attention-related pairs. During a co-creative AI-assisted conversation, children exhibit a less coherent and integrated control–attention–memory architecture than adults. The combination of similar creativity scores with divergent neural profiles suggests that developmental differences in the implementation of creative cognition may be more pronounced than baseline capacity. This highlights FP-centric pathways and their salience-gated interactions as candidate substrates by which creativity supports goal-directed co-creative dialog with AI.

Highlights

• Children demonstrate a lower within-network connectivity in the frontoparietal, salience, and dorsal attention systems while using ChatGPT compared to adults.

• Reduced between-network integration in cognitive control and modulation networks (frontoparietal–salience and frontoparietal–dorsal) in children vs adults while using ChatGPT.

• In adults, but not in children, higher creativity was associated with stronger connectivity within cognitive control networks.

Introduction

Digital tools have become deeply embedded in children’s daily lives worldwide. Research shows that these technologies, ranging from tablets and laptops to smartphones and emerging generative artificial intelligence (AI), have a multifaceted impact on children’s learning, well-being, and development (Zivan et al. 2023; Hahnefeld et al. 2025; Horowitz-Kraus et al. 2024; Nagata et al. 2024; Sundqvist et al. 2024; Jomaa et al. 2025; Wan et al. 2025). While digital tools may offer substantial opportunities to enhance educational experiences, they also introduce challenges such as distraction, cognitive overload, social–emotional strain, and negative impact on literacy development.

In many countries, fostering digital competence and inclusion is a strategic priority, with a vision that this implementation will promote a resilient, innovative, knowledge-based society, and will spark creativity. Technology can support learning, but it cannot lead it. Digital tools are not a substitute for educators; they must be used in partnership with them. Children learn best when accompanied by someone who guides, encourages, challenges, and believes in them, a feeling often evoked by generative AI tools(Chin et al. 2024). However, how much children are actually engaged cognitively and creatively when interacting with Chat-GPT is yet to be known, which is the topic of the current study.

Generative AI use by children

Children’s interactions with generative AI (GenAI) are emerging as both opportunities for creativity and challenges for critical engagement. Experimental studies have shown that tools like ChatGPT can support young children’s storytelling by providing interactive, adaptive dialog (Chin et al. 2024). For example, a study of 5–6-yr-olds found that children enjoyed storytelling with ChatGPT, which provided more frequent positive feedback and open-ended prompts than parents did, suggesting potential for supporting children’s creative expression when parental time is limited (Chin et al. 2024). Similarly, interventions using games such as Bot VoyAIge engaged 10–12-yr-olds in critically examining ChatGPT’s responses, revealing that children could identify gender, cultural, and age biases embedded in its outputs (Shrivastava et al. 2024). These findings highlight that while children can find GenAI engaging and even preferable in certain contexts, structured support is necessary to help them navigate its biases and limitations.

Beyond creative engagement, researchers emphasize the developmental, societal, and ethical implications of children’s growing exposure to GenAI. Cross-national work applying a child rights framework has underscored the need to prioritize children’s voices and experiences in shaping AI governance, pointing to rights such as privacy, safety, participation, and development that are often overlooked in current policy debates (Stoilova et al. 2025). Broader reviews caution that AI may reshape learning, attention, and social interactions, particularly during sensitive developmental windows, requiring careful regulation and proactive safeguards (Neugnot-Cerioli and Laurenty 2024). Educational interventions that build AI literacy, such as teaching middle schoolers to recognize deepfakes and misinformation, show promise in fostering critical awareness of GenAI’s socio-technical dimensions (Ali et al. 2021). Collectively, these studies demonstrate that children can be both active creators and critical evaluators of GenAI, but ensuring safe, equitable, and developmentally appropriate use will require interdisciplinary collaboration and child-centered policy.

Many fields, including the field of creativity research, are now experiencing and contending with the rise of AI models such as ChatGPT (Achiam et al. 2023) which are becoming increasingly popular tools in many linguistic Human-Computer Interactions. The emergence of publicly accessible AI models such as ChatGPT has given rise to global conversations on the implications of AI capabilities and the potential for co-creative cooperation between human and machine agents (Rafner et al. 2023; Hubert et al. 2024). Co-creativity is generally defined as multiple parties contributing to a creative process in a blended manner (Wan et al. 2024). Co-creativity is becoming especially relevant in the age of generative AI models (Rafner et al. 2023), where highly advanced generative AI can now begin to match humans in creativity tests (Hubert et al. 2024). However, increasing evidence highlights not only the benefits (de Chantal and Organisciak 2025; de Chantal et al. 2025, 2026; Kreisberg-Nitzav and Kenett 2025), but also the potential costs of AI on human creativity (Cropley 2023; Doshi and Hauser 2024; Moon et al. 2025a, 2025b; Wenger and Kenett 2025). Thus, elucidating potential differences between children and adult co-creative interactions is greatly needed, and especially elucidating the neural mechanisms that relate to such potential differences in co-creative acts. This is because children’s development continues to shape large-scale neural wiring and may be affected by the rapid AI technological immersion.

Brain development

A recent study utilizing Electroencephalogram (EEG) examined the cognitive engagement in adult individuals who were instructed to write an essay under several conditions: while using ChatGPT, search engines, or without any assistance. Another condition allowed those who used ChatGPT to write without any assistance, and those who wrote without any assistance were allowed to use ChatGPT for writing (Kosmyna et al. 2025b). EEG coherence data demonstrated a cognitive underengagement in ChatGPT users: For writing assignments, individuals were found to engage fewer cognitive resources than individuals who were using other tools to gain knowledge (such as Google) (Kosmyna et al. 2025b). Adults, however, have well-developed cognitive, attention, and modulation systems (Posner and Petersen 1990; Posner 2011; Petersen and Posner 2012). Neurobiologically, brain networks supporting these cognitive control abilities include the frontoparietal and cingulo-opercular (Executive Functions; EF system), Dorsal and Ventral Attention Networks (DAN, VAN, bottom-up attention networks), and the modulation system (Default Mode Network; DMN, and the Salience network) (Dosenbach et al. 2008), reach their full developmental maturation at the age of 26 yr (Casey et al. 2000; Giedd 2004; Lenroot and Giedd 2006). The long-range connections and the efficiency of communication between these networks increase markedly through development, reaching maturity in adulthood (Fair et al. 2012). While generating new ideas, as happens when individuals communicate or when they speak to generative AI, there is a need to plan, inhibit, and remember what was already discussed (i.e. working memory)- all parts of EF, which are not yet mature among young 6-yr-old children(Anderson 2002). However, creativity is also important for generating new ideas (Benedek et al. 2014). What are the neural mechanisms that support the involvement of EF during a co-creative task such as an interaction with ChatGPT, how do they differ in children versus adults, and how might they be impacted in relation to interaction with ChatGPT?

Creativity in children

Creativity is broadly defined as the ability to generate ideas that are both novel and useful within a given social context (Runco and Jaeger 2012; Leggett 2017; Green et al. 2024). This construct relies on cognitive processes, including divergent thinking (DT), which encompasses fluency, originality, and flexibility (Vaisarova and Carlson 2021; Bai et al. 2024), memory (Benedek et al. 2023; Kenett 2025), cognitive control (Chrysikou 2019), attention (Benedek and Fink 2019), and associative thinking (Beaty and Kenett 2023). Developmental perspectives emphasize that creativity is not a fixed trait but an emergent process, with EF providing regulatory mechanisms to support idea generation and sustain attention (Tzachrista et al. 2023).

However, empirical findings suggest that young children’s DT may rely heavily on bottom-up associative processes, with EF sometimes playing an inverse or constraining role (Vaisarova and Carlson 2021). Critically, current research focusing on the creative process emphasizes dull-process theories, mainly those including creative idea generation and idea evaluation stages (Sowden et al. 2014; Boot et al. 2017; Kleinmintz et al. 2019; Benedek et al. 2023). A growing scientific focus regards creative idea evaluation, and how it is related to metacognitive capacities, which develop later in life (Lebuda and Benedek 2023). Longitudinal studies show substantial gains in fluency and originality between ages 4 and 6, supported by both memory retrieval and mental operations, but also highlight a potential “slump” in creativity when children enter first grade and encounter more formalized, convergent instructional practices (Bai et al. 2024).

The assessment of creativity has often relied on DT tasks such as the Alternative Uses Task (AUT), which capture both product-level measures (e.g. fluency and originality) and process-level mechanisms underlying idea generation (Runco 1991). These tasks are advantageous in their sensitivity to developmental change and in revealing not only the quantity but also the quality of children’s ideas. Complementary measures of evaluative ability are also necessary, as they assess whether children can judge the originality or appropriateness of their own ideas (Runco 1991). Such current assessment tools are based on computational methods such as large language models to provide automatic and quantitative measures of originality and appropriateness of responses in open-ended creativity tasks such as the AUT (Kenett 2019; Beaty and Johnson 2021; Organisciak et al. 2023; Patterson et al. 2025).

Contemporary accounts highlight that creativity is shaped by social interaction and cultural context, positioning it as a developmental process embedded in children’s everyday environments (Sawyer et al. 2003; Leggett 2017). While studies using AI with children demonstrate that generative models can scaffold storytelling and imaginative play, the extent to which such tools enhance or constrain DT remains insufficiently understood, and whether cognitive processes are engaged equally as in adults further requires empirical work.

The neuroscience of creativity

Over the past few decades, extensive neuroscientific research has been conducted to elucidate the neural mechanisms that underlie creative thinking (Abraham 2013, Benedek et al. in press). Such research has been conducted via multiple neuroscientific methods (Benedek et al. 2019), and extensively via functional magnetic resonance imaging (MRI) research to elucidate functional brain connectivity that supports creativity (Beaty et al. 2016, 2018, 2019; Frith et al. 2021; Chen et al. 2025).

Functional magnetic resonance imaging (fMRI) has significantly advanced the understanding of creativity by highlighting how distributed brain networks support creative cognition. Recent studies have emphasized functional interactions among large-scale brain systems composed of multiple regions. This approach has identified several key networks—specifically, the EF system, and the modulation one (DMN, and salience network)-whose coordinated activity underlies various facets of creative thinking (Beaty et al. 2019).

The DMN, as part of the modulation system, comprising regions such as the medial prefrontal cortex, posterior cingulate cortex, angular gyrus, and hippocampus—has emerged as central to creative cognition due to its role in memory retrieval and associative thinking (Luchini et al. 2025b). The DMN supports memory processes essential to creative ideation, especially remote associative thinking—the ability to link conceptually distant ideas (Beaty et al. 2017, Marron et al. 2020, Beaty and Kenett 2023). Beyond idea generation, the DMN also contributes to evaluating ideas by comparing them to stored knowledge to determine novelty and by mentally simulating their feasibility to assess appropriateness (Benedek et al. 2023). The salience network is comprised of the anterior insula and dorsal anterior cingulate cortex—orchestrates interactions between the DMN and the EF system, dynamically transitioning between generative and evaluative modes (Beaty et al. 2016, 2019; Chen et al. 2025). Supporting this gating function, multilayer network analyses demonstrate extensive reconfiguration among the EF system, DMN, and SN during creative idea generation, while evaluation engages a more integrated yet stable network architecture (Kenett et al. 2024).

Alongside the modulation system, the EF networks—comprising primarily lateral prefrontal and posterior parietal cortices—provides top-down regulation to steer creative thinking toward task goals, such as response inhibition and idea revision that are critical for refining creative ideas. The EF system is particularly involved in suppressing dominant but less original responses, enabling individuals to consider novel, more creative alternatives (Beaty et al. 2017). Such cognitive control engagement tends to vary with creative task demands (Chrysikou et al. 2014), with flexible down-regulation often facilitating originality when dominant responses must be overcome (Chrysikou 2019). Creative thinking is characterized by heightened cooperation among these networks, with stronger coupling between the DMN and EF consistently predicting higher creative performance (Beaty et al. 2016, 2019). Connectome-based predictive modeling reliably links enhanced functional connectivity across DMN, EF, and SN regions to individual differences in creative ability (Beaty et al. 2018, Frith et al. 2021, Ovando-Tellez et al. 2022). Highly creative individuals exhibit greater network flexibility, showing more frequent transitions between the DMN and EF system (Chen et al. 2025) and increased temporal variability, especially within the DMN (Sun et al. 2019). How critical is the involvement of these cognitive control networks for the realization of creativity potentially impacted by interaction with ChatGPT, and even more differentiated in children and adults in the question of the current study.

Why first grade? First grade (age 6–7) marks a developmentally distinct window for studying children’s engagement with generative AI. This period coincides with the transition from pre- to formal literacy instruction, when children are first learning to independently extract and evaluate information from text and speech. First graders also show characteristically high trust in adults and technological authorities, tending to accept information from a perceived expert source with less independent scrutiny than older children(Girouard-Hallam and Danovitch 2022, Flanagan et al. 2023). At the same time, EF and theory of mind are undergoing particularly rapid development during this period(Benson et al. 2013, Toker and Acarlar 2026) making first grade a sensitive window in which the cognitive-control demands of a novel, adaptive conversational partner such as ChatGPT may be especially pronounced. Understanding how children in this age group engage cognitively with generative AI is therefore critical both for informing developmentally appropriate use in early education and for identifying the neurocognitive mechanisms most vulnerable to being under- or over-taxed by such tools.

The current study

The current study aimed to determine the differences in cognitive engagement when children and adults are engaged while communicating with ChatGPT during a co-creative task. The study also examined whether the level of creativity was related to the level of cognitive control engagement. We hypothesized that children would show a lower engagement of cognitive control networks and that a connection between creativity and cognitive control functional connectivity would be observed in adults, but not in children.

Methods

Participants

Fifteen 6–7-yr-old children, all first graders (mean age: 6.44 yr, SD = 4 mo 11 males) and 16 adults (mean age: 38.12 yr, SD = 7.25 yr, 4 males) from an average-above-average socioeconomic status participated in the current study. For both age groups, the following inclusion criteria were applied: (1) intact vision and hearing, (2) no history of psychiatric or neurological delays (3) intact non-verbal and verbal IQ (>85), (4) born term (>37 wk gestational age), (5) no contradiction to MR scanning: claustrophobia, any metal object in the body, including cardiac pacemaker, dental braces, and cochlear implants. Participants were recruited via [insert actual channel(s), e.g. the Educational Neuroimaging Group’s existing participant registry, flyers distributed through the local community and schools, and social media outreach]. Socioeconomic status was estimated from (insert basis, e.g. parental education and the recruitment area’s socioeconomic ranking), which corresponded to an average-to-above-average range for the general population.

Participants underwent a comprehensive behavioral assessment of their EF and creativity abilities, as well as an fMRI scan. Children and their parents signed informed assent and consent forms, respectively, and adults signed written consents before participating. Data were collected at the Technion Human MRI Research Center (TechMRC). The protocol was approved by the Ministry of Health Helsinki Committee and Schneider’s Hospital Medical Center (protocol number: 0531-16-RMC, for children) and by the institutions’ ethical committee at the Technion (protocol number: 179-2023). Participants were compensated for their participation.

Study procedure

Once children and adults were recruited and consented to participate in the study, they were invited to the Educational Neuroimaging Group (ENIG) facility at the Technion, where their cognitive and creative abilities were assessed. Then, upon completion, families were invited to the MRI facilities at the Technion, where ChatGPT was set up on the MRI console, and the child’s voice and all communication were recorded. All conversations with ChatGPT were recorded and analyzed for semantic complexity, a factor previously linked to creativity [following (Raz et al. 2024)]. The children-versus-adults comparison was chosen to isolate developmental differences in cognitive-control engagement under an identical co-creative task demand, rather than to compare engagement with versus without ChatGPT; a no-ChatGPT control condition was not included in either group.

Behavioral measures

To verify intact nonverbal IQ, children completed the Test of Nonverbal Intelligence [TONI (Brown et al. 2010)] which includes a completion of missing parts in given matrices. To evaluate language abilities, the vocabulary test from the Peabody Picture Vocabulary Test [PPVT (Dunn and Dunn 2007)] was used. For adults, equivalent tests were administered: the Matrix task and the vocabulary task from the Wechsler Adult Intelligence Scale (WAIS) (Wechsler 2008).

Creativity was assessed using the divergent association task (DAT (Olson et al. 2021), https://www.datcreativity.com/task), conducted outside of the scanner. In this task, participants were required to retrieve 10 words that are as unrelated as possible, without any time constraints. The semantic relations between the words were calculated with a higher value representing higher creativity abilities (more divergent thinking).

Neuroimaging measures

Talking to ChatGPT paradigm

The fMRI data were collected using a Prisma Magneton 3 Tesla scanner (Siemens Healthineers, Erlangen, Germany). A unique infrastructure set-up at the TechMRC allowed the data collection during the communication wth ChatGPT: an internet connection to the audio-visual system in the scanner, a real-life speech recording while fMRI data is collected using Optoacoustics system (Optoacoustics Ltd, Mazor, Israel), an in-house noise reduction system developed at the Technion allowing a clear sound transmission of the ChatGPT audio stimuli, and another in-house developed algorithm which allows a noise elimination from the audio recording allowing a direct transcript of the verbal output of the child. After desensitization and acclimation to the scanner environment [following our previous experience in this age group (Kraus and Horowitz-Kraus 2022)], anatomical data (T1) was collected for coregistration purposes. Upon connecting to ChatGPT, children were informed that they were about to speak to the “smartest computer in the world,” also known as ChatGPT, whereas adults were simply told that they would be speaking to ChatGPT (see prompts in the Supplemental Materials). Both groups were told that their goal is to open a store; It can be any store they would like, it can be anywhere, for anyone. They need to consult with ChatGPT regarding the type of store for 5 min while the child speaks to ChatGPT. All participants, children and adults alike, were explicitly instructed to respond verbally throughout the interaction. Note, the instruction was given only when they were in the scanner, immediately before the start of the scan. An audio recording of the entire conversation was conducted via Audacity software (https://www.audacityteam.org/) and the transcript of the conversation was generated using ChatGPT.

Neuroimaging data

Each fMRI session lasted 5 min exactly and included a Time Repetition (TR) of 1 sec and the echo time (TE) was 28 ms. The imaging setup included a field of view (FOV) of 20 × 20 × 13.6 cm, with a 100 × 100 matrix and slices of 2 mm thickness. High-resolution T1 images were collected for each participant to align the functional images accurately. The T1 scan parameters were a TR of 1750 ms, a TE of 2.8 ms, an inversion time of 900 ms, and a flip angle of 9°, with a field of view (FOV) of 22.4 × 22.4 × 16 cm, a 224 × 224 × 160 matrix, and a slice thickness of 1 mm. All scans included a total of 300 volumes per session.

Neuroimaging data processing

The fMRI data were corrected for geometrical distortion and signal intensity variation due to static B0 field variations in T2^* weighted fMRI data using corresponding B0 field map images acquired in every participant. Rigid boundary-based registration (Greve and Fischl 2009) to a high-resolution 3D, T1-weighted anatomical volume acquired within the same session, was performed. Nonlinear warping of this anatomical volume into standard space using a pediatric template was performed using ANTs (Avants et al. 2008, Klein et al. 2009). Functional MRI data were subjected to the same preprocessing strategy outlined in (Horowitz-Kraus 2015, Levinson et al. 2018, Twait 2018, Farah et al. 2019, Farah and Horowitz-Kraus 2019, Meri et al. 2020, Taran et al. 2022). Using a model-based approach, we constructed a model for the functional connections among brain regions involved in the reading network (Turkeltaub et al. 2003, Karunanayaka et al. 2011, Twomey et al. 2011, Richlan 2012, Price 2013, Vogel et al. 2013, Freedman et al. 2020). This included the export of functional connectivity values for the selected networks, as well as EF, attention and modulation networks (CO, FP, memory, DAN, VAN, DMN, and Salience [Vogel et al. 2014]) using the CONN toolbox implemented in SPM 12 ([Whitfield-Gabrieli and Nieto-Castanon 2012]; http://www.nitrc.org/projects/conn/). Functional connectivity between brain regions was quantified by the temporal correlation between their spontaneous BOLD signal fluctuations (Power et al. 2014) based on Power’s atlas (Power 2010) containing 264 putative functional parcels, covering cognitive, sensory, and language/auditory functional networks (see Fig. 1 for the networks’ spatial masks). The Power atlas was selected because it remains one of the most extensively validated and widely applied functional parcellations in the developmental and adult fMRI literature, supporting comparability with prior work. Because functional data were first nonlinearly warped into a pediatric template (as described above) prior to parcel-wise extraction, anatomical misalignment between the adult-derived parcellation and children’s brains was minimized.

Figure 1.

Executive functions, attention, and modulation networks’ masks.

To control for motion artifacts, functional connectivity analysis included a rigorous quality control (QC) process, which involved calculating the mean framewise displacement (FD) by summing the absolute values of the derivatives of the translational and rotational realignment estimates (after converting the rotational estimates to displacement at a 50 mm radius). The QC vector was correlated across participants and conditions with the pairwise functional connectivity correlations between all parcels to ensure that motion artifacts do not drive the connectivity [see (Power et al. 2015)]. None of the volumes reached FD > 0.5 mm, and hence segments were not removed from the data [per the acceptable definition of data lasting fewer than five contiguous volumes were flagged and removed; (Gordon et al. 2016, Power et al. 2015)]. In addition, motion parameters were included as covariates in the first-level analyses. Due to significant differences in FD in children (X = .306, SD = .06) and adults (X = .226, SD = .05), t(30) = 3.65, P < .001), FD was controlled for in all analyses involving neuroimaging data.

Divergent sematic integration generated from the conversations with ChatGPT

Divergent sematic integration (DSI) calculates the average semantic distance between all word pairs in a response using embeddings from a pre-trained model and cosine similarity. DSI has been shown to significantly predict originality on other creative tasks involving long-form text responses (Johnson et al. 2023, Luchini et al. 2025a). Semantic-distance models work by placing words, sentences, or questions into a shared “semantic space,” where the distance between them reflects how similar or different their meanings are (Kenett 2019). For example, cat and dog will appear closer together in this space because they share many contextual similarities, while cat and democracy will be much farther apart. This is done in an unsupervised manner meaning they learn patterns of word co-occurrence and meaning directly from large text corpora without relying on predefined human ratings or labeled training data (Beaty and Kenett 2023). DSI scores were computed for each participant for their entire discussion with ChatGPT based on their narrative transcripts. This was done to ensure enough text in each narrative to reliably compute DSI score. Thus, each DSI score represents the general diversity, or richness, of the conversation with ChatGPT. A higher score denotes a richer, more original narrative interaction with ChatGPT.

Statistical analyses

Behavioral data analysis

Several t-test analyses were conducted to determine the differences in creativity levels assessed inside (DSI) and outside (DAT) of the scanner, among children and adults.

Neuroimaging data analysis

To determine differences in within- and between-network functional connectivity between children and adults, a multivariate analysis of variance (MANOVA) was conducted with group (children, adults) as a between-subjects factor and framewise displacement (FD) as a covariate to control for head motion. Dependent variables were mean Fisher-z-transformed connectivity values within seven canonical large-scale brain networks: cingulo-opercular (CO), default mode (DMN), memory retrieval, frontoparietal (FP), salience, ventral attention (VAN), and dorsal attention (DAN) networks, as well as their combinations. Univariate tests followed significant group effects and post hoc pairwise comparisons. Partial eta-squared (η²_P) values were reported as effect sizes.

Moderation analysis for the relationship between neuroimaging and behavioral data (creativity percentile scores)

To examine whether creativity moderates group differences in functional connectivity, we extended the above models by adding the creativity percentile score as a covariate and including the group-by-creativity interaction term.

Separate multivariate General linear model (GLM) models were estimated for both within- and between-network connectivity: (1) Model predictors: group (children vs. adults), creativity (centered), FD, and the group × creativity interaction, and (2) Dependent variables: within- or between-network connectivity measures as described above. For significant interactions, simple slope analyses and predicted values were examined to probe the direction and nature of the moderation effects.

Results

Behavioral results

Both children and adults demonstrated average verbal and nonverbal general abilities, with no significant differences between adults and children in these measures, as determined by independent-samples t-tests. Group comparisons revealed no significant difference in creativity performance (DAT) outside of the scanner between children and adults. However, creative interaction with ChatGPT (DSI) were significantly higher in adults compared to children (Table 1).

Table 1.

Comparisons between children and adults in general abilities and creativity scores.

		Groups		t-test
Domain	Measure	Children	Adults
		Mean(SD)	Mean(SD)
Nonverbal abilities	Matrices	10.41 (1.99)	12.93 (2.2)	−3.354, P < 0.01
Verbal abilities	Vocabulary	11.2 (3.09)	13.92 (3.25)	−2.72, P < 0.05
Creativity	DAT	28.61 (19.59)	26.16 (21.53)	0.33, NS
	DSI (while speaking to ChatGPT)	.81 (.01)	.82 (.005)	−3.67, P < 0.001

Note. SS = Standard Score (M = 100, SD = 15), ScS = Scaled Score (M = 10, SD = 3) ^*  P < 0.05, ^**  P < 0.01, ^***  P < 0.001, n.s. = non significant.

Neuroimaging measures

Within networks analysis: After statistically controlling for head motion (FD), significant group differences in functional connectivity were observed in the FP network, F(1,27) = 8.77, P = 0.006, partial η² = .245, the salience network, F(1,27) = 5.65, P = 0.025, partial η² = .173, and the DAN, F(1,27) = 5.49, P = 0.027, partial η² = .169, with adults showing higher FC within these networks compared to children. These effects correspond to medium-to-large effect sizes, indicating that a substantial proportion of variance in connectivity was explained by group differences even after adjusting for motion.

Between Networks analysis: After statistically controlling for head motion (FD), several significant group differences emerged in between-network functional connectivity (Table 3). Compared with children, adults exhibited significantly stronger connectivity between the CO-FP networks, F(1,27) = 5.628, P = 0.025, η²_P = .173) and between the CO-Salience networks (F(1,27) = 8.182, P = 0.008, η²_P = .233. Similarly, adults showed greater coupling between the DMN-Memory, F(1,27) = 6.491, P = 0.017, η²_P = .194, and DMN-FP, F(1,27) = 5.570, P = 0.026, η²_P = .171. After correcting for multiple comparisons (using FDR correction (Benjamini and Hochberg 1995), Benjamini–Hochberg, q = .05), the results for the FP findings remained significant. See Table 2 for the networks results.

Table 3.

Between networks, functional connectivity of EF, attention, and sensory after controlling for FD.

Network Pair	Children (Mean ± SD)	Adults (Mean ± SD)	ANCOVA Results (F(1,27), p, η 2  P)	Contrast
CO–DMN	0.133 ± 0.070	0.119 ± 0.123	F = 0.006, P = 0.937, η2P = .000	C > A
CO-Memory	0.015 ± 0.096	0.085 ± 0.110	F = 3.453, P = 0.074, η2P = .113	A > C
CO–FP	0.043 ± 0.074	0.105 ± 0.052	F = 5.628, P = 0.025, η  2P = .173	A > C
CO-Salience	0.109 ± 0.076	0.207 ± 0.079	F = 8.182, P = 0.008, η  2P = .233	A > C
CO-VAN	0.165 ± 0.123	0.182 ± 0.103	F = 0.311, P = 0.582, η2P = .011	A > C
CO-DAN	0.121 ± 0.077	0.116 ± 0.086	F = 0.494, P = 0.488, η2P = .018	C > A
DMN–Memory	0.111 ± 0.045	0.150 ± 0.091	F = 6.491, P = 0.017, η  2P = .194	A > C
DMN – FP	0.090 ± 0.039	0.115 ± 0.065	F = 5.570, P = 0.026, η  2P = .171	A > C
DMN–Salience	0.111 ± 0.060	0.115 ± 0.081	F = 0.480, P = 0.494, η2P = .017	A > C
DMN–VAN	0.196 ± 0.072	0.188 ± 0.091	F = 0.015, P = 0.902, η2P = .001	C > A
DMN–DAN	0.116 ± 0.073	0.118 ± 0.086	F = 0.896, P = 0.352, η2P = .032	A > C
Memory–FP	0.079 ± 0.064	0.104 ± 0.096	F = 6.449, P = 0.017, η  2P = .193	A > C
Memory-Salience	0.049 ± 0.059	0.079 ± 0.103	F = 3.990, P = 0.056, η2P = .129	A > C
Memory-VAN	0.091 ± 0.073	0.133 ± 0.152	F = 5.514, P = 0.026, η  2P = .170	A > C
Memory-DAN	0.086 ± 0.076	0.144 ± 0.100	F = 5.944, P = 0.022, η  2P = .180	A > C
FP–Salience	0.138 ± 0.040	0.180 ± 0.060	F = 2.434, P = 0.130, η2P = .083	A > C
FP–VAN	0.095 ± 0.059	0.167 ± 0.088	F = 5.820, P = 0.023, η  2P = .177	A > C
FP–DAN	0.039 ± 0.059	0.126 ± 0.043	F = 13.252, P < 0.001, η  2P = .329	A > C
Salience–VAN	0.012 ± 0.000	0.189 ± 0.007	F = 1.687, P = 0.205, η2P = .059	A > C
Salience–DAN	0.010 ± 0.005	0.145 ± 0.005	F = 1.824, P = 0.188, η2P = .063	A > C
VAN–DAN	0.002 ± 0.007	0.186 ± 0.007	F = 0.338, P = 0.566, η2P = .012	A > C

VAN: Ventral attention network; DAN: Dorsal attention network; FP: Fronto-parietal; CO: Cingulo-opercular; DMN: default mode network; C: children; A: adults. Bold represents significant results.

Table 2.

Within networks functional connectivity of EF, attention, and sensory networks.

Network	Children (Mean ± SD)	Adults (Mean ± SD)	ANCOVA Results (F(1,27), p, partial η 2 )	Contrast
Cingulo-Opercular	0.307 ± 0.128	0.306 ± 0.106	F = 0.006, P = 0.940, partial η2 = .000	C ≈ A
Frontoparietal (FP)	0.113 ± 0.036	0.201 ± 0.085	F = 8.773, P = 0.006, partial η  2 = .245	A > C
Memory	0.265 ± 0.168	0.342 ± 0.124	F = 1.938, P = 0.175, partial η2 = .067	A > C
Default Mode (DMN)	0.181 ± 0.054	0.187 ± 0.057	F = 0.497, P = 0.487, partial η2 = .018	A > C
Salience	0.198 ± 0.073	0.289 ± 0.078	F = 5.653, P = 0.025, partial η2 = .173	A > C
Dorsal Attention (DAN)	0.162 ± 0.098	0.211 ± 0.086	F = 5.487, P = 0.027, partial η2 = .169	A > C
Ventral Attention (VAN)	0.369 ± 0.107	0.344 ± 0.120	F = 1.562, P = 0.222, partial η2 = .055	C > A

VAN: Ventral attention network; DAN: Dorsal attention network; FP: Fronto-parietal; CO: Cingulo-opercular; DMN: default mode network; C: children; A: adults. Bold represents significant results.

Enhanced between networks connectivity in adults was also observed between the Memory Retrieval network and multiple systems, including the FP, F(1,27) = 6.449, P = 0.017, η²_P = .193, VAN, F(1,27) = 5.514, P = 0.026, η²_P = .170, and DAN F(1,27) = 5.944, P = 0.022, η²_P = .180. Finally, significant group effects were found for connectivity between the FP-VAN, F(1,27) = 5.820, P = 0.023, η²_P = .177, and FP-DAN, F(1,27) = 13.252, P = 0.001, η²_P = .329, with adults again showing higher connectivity.

Overall, these results suggest a broader and more integrated pattern of between-networks connectivity in adults relative to children, particularly involving the FP, DAN,VAN, memory, and salience systems. After correcting for multiple comparisons (FDR correction (Benjamini and Hochberg 1995), Benjamini–Hochberg, q = .05), the results for the FP-salience and FP-DAN findings remained significant (Table 3). See also Fig. 2.

Figure 2.

Averages for within and between networks’ functional connectivity (Z scores). The Z-score averages for within and between networks functional connectivity for children (A), adults (B) and the difference between the groups (C). The hot color represents a higher Z values whrease the cool color represents lower Z values.

Moderation results

In order to determine whether creativity scores moderated the group effect on within-network connectivity (controlling for FD), we conducted moderation analyses, which revealed no significant group × creativity interaction effects across any network (P > 0.05).

Moderation analyses revealed that creativity outside the scanner (DAT) significantly interacted with group in predicting between-network functional connectivity. Specifically, significant moderation effects emerged for the FP–Salience, F(1,26) = 5.869, P = 0.023, partial η²_P = .184, and Salience–VAN, F(1,26) = 4.575, P = 0.042, partial η²_P = .150, pairs, indicating that the relationship between creativity and connectivity strength differed between children and adults for these networks.

To determine which group (children vs adults) shows a greater association between creativity skills (DAT) and functional connectivity between FP-Salience networks, an indepth slope analysis for the significant interactions was conducted, and revealed that adults’ show a positive significant slope (slope B = +0.028 per 1 SD creativity, P = 0.022, 95% CI [0.004, 0.052]) whrease children demonstrate a negative non significant slope (B = −0.014, P = 0.293, 95% CI [−0.042, 0.013] (n.s.)). Similar directions were found for Salience–VAN networks as well (Adults: B = +0.037, P = 0.062, 95% CI [−0.002, 0.077]; Children: B = −0.025, P = 0.273, 95% CI [−0.070, 0.021] (n.s.)). Overall, higher creativity is associated with stronger FP–Salience coupling in adults, but not in children (Fig. 3).

Figure 3.

Association between creativity, functional connectivity and group differences. (A) Correlation analysis between the functional networks and creativity score. (B) Slope differences analysis, comparing children to adults.

Power analysis

Power calculations were anchored to published between-group differences in large-scale network connectivity. In younger versus older adults, Avelar-Pereira et al. reported a robust task-state difference in left FP-DAN coupling during the Multi-Source Interference Task, with groups of 29 and 30 and a large effect (Cohen’s d ≈ 1.3) (Avelar-Pereira et al. 2017). In a separate high-versus low-creative adult comparison (12 per group), Beaty et al. reported ROI-to-ROI group differences spanning t values of 2.38–3.75 for IFG–DMN links, corresponding to large effects (d ≈ 1.0–1.5) (Beaty et al. 2014). To adopt a conservative planning assumption, we powered a priori for a moderate effect (Cohen’s d = 0.60–0.70), which requires ~44–33 participants per group for 80% power at α = .05 (two-sided). Our achieved sample (children, n = 15; adults, n = 16) is therefore well-powered for large effects, consistent with prior literature, but underpowered for moderate effects, guiding the interpretation of null or small effects accordingly.

Discussion

How old do you need to be to use generative AI? With the invasion of these tools into our lives and even into the educational system (Sahu 2024), the current study aimed to explore the differences in cognitive control engagement when children vs. adults were orally communicating with ChatGPT during a creative task. In line with our hypothesis, children showed a lower engagement of cognitive control networks while communicating with ChatGPT compared to adults, despite an equal creativity level (age normalized) across groups. Moreover, and in line with our results, positive relations between creativity levels and cognitive control functional connectivity strength were observed in adults, but not in children.

Reduced cognitive control, attention, and modulation networks during conversations with GenAI tools in children

Talking to ChatGPT or any other generative AI tools provides a unique opportunity to quickly gather information and build upon it to generate more higher-level, creative ideas. Whereas knowledge was previously gathered using search engines or other printed paper sources, today, gathering basic information is faster, thanks to the efficiency prompts provided by the user(Huang and Huang 2024). This process demands planning the utterance, monitoring your response and ChatGPT’s response, reorienting one’s attention to unpredictable responses, and maybe inhibiting responses when the conversation takes a different direction—these abilities all fall under the EF system definition (Cirino 2023). Whereas these stages may not dramatically differ in children and adults during human-to-human interaction, the speed of information flow and responses from ChatGPT and the lack of gestures usually received from humans but equipped with other verbal gestures(Welivita and Pu 2024), might make this task very different than communicating with humans.

It was previously suggested that using generative AI tools results in lower alpha and beta waves generated from EEG data, with the claim that these tools demand less cognitive effort than an actual active search (engines vs self-search with no external tools) (Kosmyna et al. 2025a). The current study suggests that where adults do show an engagement of cognitive control (FP), attention (DAN), and modulation (salience) systems while speaking to ChatGPT, children show reduced functional connections within these networks. These findings are in line with developmental theories previously suggested (Jolles et al. 2011).

Children also demonstrated reduced functional connectivity between the selected cognitive networks (FP-salience, FP-DAN, DMN-memory). In line with the reported increased between-networks connectivity during development (Fair et al. 2012), it appears that, when communicating with ChatGPT, children are less capable of retrieving information and internal mentation (DMN-memory), or adapting the behavior/thoughts in response to changes in a given stimulus (FP-DAN). These results, combined with studies demonstrating a lack of maturation of the cognitive control system at the age of 6 yr (Anderson 2002) (first graders, the age of the participants in the current study), suggest that speaking to ChatGPT, which seems not to be so complicated, actually demands cognitive control, which is lower in children who participated in our study. This is especially true as the children in the current study are in their first grade, and there is a debate regarding the most appropriate age for using ChatGPT in school. Additional studies are needed to determine the exact time window in development when children gain the most from interacting with ChatGPT, also in terms of cognitive demands.

Differential engagement of cognitive control networks in creative thinking: children versus adults

Creativity, measured in this study as DAT (Olson et al. 2021), is related to cognitive control, with evidence demonstrating increased creative abilities with increased frontal activation (specifically the Dorsolateral prefrontal cortex) from adolescence to young adulthood (Jolles et al. 2011). Interestingly, although children and adults in the current study exhibited comparable levels of creativity measured outside of the scanner (age-normalized DAT), only adults showed significant associations between functional connectivity among cognitive control, attention, and modulation networks (FP–salience, salience–VAN) networks involved in the switching of EFs and attention (Schimmelpfennig et al. 2023) critcal for creative thinking (Beaty et al. 2015, 2018, 2019, Kenett et al. 2024, Chen et al. 2025). In contrast, no such relations were observed in children.

Interestingly, previous research has linked creative thinking to the involvement of the salience network, which identifies and filters important or novel ideas before relaying them to the executive control system for goal-directed processing (Beaty et al. 2015, 2018, 2019). Notably, the DMN, proposed as a key component in models of creativity (Shofty et al. 2022; Luchini et al. 2025b) that support spontaneous thought and flexible memory retrieval, did not differ between adults and children, either overall or in relation to their DAT scores. This lack of difference may reflect the novelty of interacting with ChatGPT for both groups.

An additional possible explanation for these findings between-networks connectivity results and the absence of relations to creativity measures in children might be that children still do not have the neurobiological infrastructure to benefit from the conversation with ChatGPT, which also aligns with the lower DSI scores from their generated ChatGPT script. This is critical given the extended attention given to co-creative human-AI interaction (Rafner et al. 2023), and their potential costs (Doshi and Hauser 2024, Moon et al. 2025a, 2025b, Wenger and Kenett 2025) and benefits (de Chantal and Organisciak 2025, de Chantal et al. 2025, 2026, Kreisberg-Nitzav and Kenett 2025) on human creativity. Future studies should examine the interaction between familiarity with ChatGPT and DMN involvement to determine whether the absence of engagement of DMN in relation to creativity is due to the equal novelty shared among children and adults using ChatGPT.

Limitation

The results of the current study should be considered in light of the following limitations. First, the instructions for both age groups were slightly different: children were told they’d speak to “the smartest computer in the world,” whereas adults were simply told “ChatGPT.” This framing difference may itself have shaped children’s engagement style: children are known to rely heavily on the judgment of perceived experts, deferring to that judgment even when it is flawed (Williams and Danovitch 2024). Describing ChatGPT as “the smartest computer in the world” may therefore have encouraged children toward greater deference and less independent, critical evaluation of its responses than adults exhibited, potentially dampening the evaluative component of cognitive-control engagement measured here rather than reflecting only a capacity-based developmental limitation. Future studies should standardize the framing of the AI system across age groups, or manipulate perceived authority/expertise as an experimental variable to directly test its contribution to children’s cognitive engagement.

In addition, the speech behavior, i.e. the conversation features such as turn count, words/min, response latency, were not accounted for and might also drive the results of the study (i.e. higher engagement and back and forth, might explain the greater cognitive control engagement). Moreover, there are apparent differences in vocabulary knowledge that change across the lifespan (Cosgrove et al. 2023, Li et al. 2024). Such a difference, which was not controlled here, might impact the co-creative interaction within the scanner and provide an alternative explanation for our results. Future studies are needed to replicate and extend our results, for example, by controlling for vocabulary knowledge in a similar study design. Finally, Post-hoc power analysis indicated 80% power to detect medium-to-large effects (Cohen’s d ≥ 0.73) for primary between-group comparisons. The moderation analyses were exploratory and likely underpowered for interaction effects, which typically require larger samples. Future studies with a sample size of ~50 children per group are warranted to adequately test creativity × group interactions.

Relatedly, because both groups completed the same ChatGPT-based task and no no-ChatGPT (or human-partner) control condition was included, the present design cannot fully separate a general effect of development on functional connectivity during this type of task from an effect specific to interacting with a generative AI system; future studies incorporating a within-subject ChatGPT-versus-no-ChatGPT or ChatGPT-versus-human-partner comparison are needed to isolate the AI-specific contribution.

Moreover, prior familiarity with ChatGPT was not formally assessed in this study due to the lack of standard tools assessing this relatively new technological aspect. Given generational differences in technology exposure, adults likely had substantially more prior experience with ChatGPT than the first-grade children in our sample, which is a limitation of the current design and should be measured directly in future work. Lastly, we relied on an adult-derived functional parcellation (Power atlas); although data were normalized to a pediatric template prior to parcellation, a child-specific or developmentally derived atlas may better capture age-specific network topology and would be a valuable direction for replication.

Conclusions

This novel study examined the level of cognitive engagement of children while speaking to ChatGPT compared to adults, as well as how this cognitive engagement is related to creativity. The results point to the overengagement of the self-regulated networks in 6-yr-old children, which warrants additional longitudinal studies to determine the turning point where cognitive control is more engaged and that creative thinking is related to cognitive control networks, similarly to adults. Given the rapid emergence of AI in our daily lives, alongside increased discussion on the costs and benefits of AI on creativity, our results highlight the significance of between-network functional coupling during human-AI creative acts. Decision makers, especially in the education system, as well as those related to children’s health (pediatricians), should pay particular attention to the official age when children can use ChatGPT, similarly to the attention given to screen exposure.

Acknowledgments

The authors would like to thank the children, the parents, and the students participating in the current project. The authors would also like to thank the research team and students at the Educational Neueoimahging Group (of 2025) for helping with data collection and tasks’ administration. The authors would also like to thank Tuval Raz for his help in analyzing the transcripts of participants’ conversations with ChatGPT in the scanner.

Author contributions

Tzipi Horowitz-Kraus (Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing—original draft, Writing—review & editing), Yoed Kenett (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Supervision, Validation, Visualization, Writing—original draft, Writing—review & editing), Dafna Link-Sourani (Conceptualization, Data curation, Investigation, Methodology, Resources, Writing—original draft, Writing—review & editing), Emeel Ashqar (Conceptualization, Data curation, Investigation, Project administration, Writing—original draft, Writing—review & editing), and Rola Farah (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing—original draft, Writing—review & editing)

Funding

No funding to declare.

Conflicts of interest

The authors declare no conflict of interest.