Lower engagement of cognitive control in children on ChatGPT

Synthesis: Horowitz-Kraus, Kenett, Link-Sourani, Ashqar and Farah (2026) compared fifteen 6–7-year-old children with sixteen adults while each group held a naturalistic five-minute spoken conversation with ChatGPT inside an MRI scanner, given the identical task (design a store, of any kind, for anyone). Creativity measured outside the scanner (Divergent Association Task) did not differ between groups, and creative interaction with ChatGPT (Divergent Semantic Integration) was significantly higher in adults. On the neural side, adults showed stronger within-network functional connectivity in the frontoparietal control network (F(1,27) = 8.77, P = 0.006, partial η² = .245), the salience network, and the dorsal attention network, and stronger cross-network coupling — most robustly frontoparietal–salience and frontoparietal–dorsal attention, both surviving FDR correction. Higher creativity predicted stronger frontoparietal–salience coupling in adults but not children. The striking result is not a capacity gap: matched creativity scores with divergent neural profiles suggests the difference lies in how creative cognition is implemented during AI-assisted dialog — children's control–attention–memory architecture is less coherent and integrated when co-creating with ChatGPT. Tool vintage: ChatGPT (model version unspecified), sessions ran in the study window ending 2026. The authors frame this as evidence relevant to the debate over a minimum age for ChatGPT use in schools — but with one framing confound (children were told ChatGPT was "the smartest computer in the world"), no no-ChatGPT control condition, and n = 31 total, the finding is a first demonstration, not a settled verdict.

Summary

Digital tools are now embedded in children's daily lives, and generative AI has moved from search engines to conversational partners: ChatGPT's adaptive dialog can support young children's storytelling (Chin et al. 2024) and even prove more engaging than a parent's responses in some contexts. But what happens cognitively when a child talks to an Large Language Models (LLMs) has barely been examined. An earlier EEG study of adults writing essays (Kosmyna et al. 2025) found cognitive under-engagement in ChatGPT users compared with search-engine users, and executive functions (EF) — planning, inhibition, working memory, attentional reorienting — are exactly what a five-minute co-creative conversation demands, and exactly what a 6-year-old's brain has not yet matured (the frontoparietal and cingulo-opercular systems reach maturity around age 26). Creativity research adds the second half of the question: creative thinking relies on EF, attention, memory and associative processes, and in adults higher creativity is associated with stronger connectivity within cognitive-control networks. This study asked how cognitive-control engagement during ChatGPT dialog differs between first graders and adults, and how it relates to measured creativity.

The design isolated developmental differences under an identical co-creative task demand: children (n = 15, mean age 6.44 yr, 11 males) and adults (n = 16, mean age 38.12 yr, 4 males) of average-to-above socioeconomic status, all with intact nonverbal and verbal IQ (> 85), held a five-minute spoken conversation with ChatGPT inside a 3T Siemens Prisma scanner (Technion Human MRI Research Center). A custom infrastructure allowed live internet-connected audio in the scanner, real-life speech recording during scanning (Optoacoustics), in-house noise reduction, and a noise-elimination algorithm producing a clean transcript. Both groups were given the same open-ended task — "open a store; it can be any store you would like, it can be anywhere, for anyone" — and instructed to respond verbally throughout. Head-motion (framewise displacement) was significantly higher in children (M = .306) than adults (M = .226, t(30) = 3.65, P < .001) and was controlled as a covariate in every imaging analysis. Creativity was assessed twice: outside the scanner via the Divergent Association Task (retrieve 10 maximally unrelated words), and inside via Divergent Semantic Integration, computed from the semantic diversity of the ChatGPT conversation transcripts. Functional connectivity was computed within and between seven large-scale networks: cingulo-opercular, frontoparietal (FP), default mode (DMN), memory retrieval, salience, ventral attention (VAN), and dorsal attention (DAN). The study is explicitly a children-versus-adults comparison — there was no no-ChatGPT control condition in either group.

Key Findings

• Adults showed stronger within-network connectivity in control and attention systems; children did not differ only on FP. After controlling for head motion, adults had significantly higher within-network functional connectivity than children in the frontoparietal 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 dorsal attention network (F(1,27) = 5.49, P = 0.027, partial η² = .169) — medium-to-large effect sizes. Only the FP finding survived FDR correction, but the direction was consistent across the control/attention cluster.
• Cross-network integration was broader in adults, most robustly FP–salience and FP–DAN. Adults showed stronger coupling on nearly every between-network pair tested — cingulo-opercular–FP, cingulo-opercular–salience, DMN–memory, DMN–FP, memory–FP, memory–VAN, memory–DAN, FP–VAN, FP–DAN (F(1,27) = 13.252, P = 0.001, partial η² = .329, the largest effect in the study). The FP–salience (F(1,27) = 8.182, P = 0.008, partial η² = .233) and FP–DAN pairs survived FDR correction. Children's architecture during ChatGPT dialog was "less coherent and integrated" across control, attention and memory systems.
• Creativity scores matched; neural profiles did not. Divergent Association Task scores (outside the scanner) showed no significant group difference, and both groups showed average verbal and nonverbal abilities. Yet Divergent Semantic Integration of the ChatGPT conversation — semantic diversity of the actual dialog — was significantly higher in adults. Equal measured creativity coexisting with divergent neural profiles suggests developmental differences in the implementation of creative cognition during AI-assisted dialog are larger than baseline capacity differences.
• The creativity–connectivity relationship existed only in adults. Moderation analyses found no significant group × creativity interaction for within-network connectivity, but a significant interaction for between-network connectivity: in adults, higher creativity predicted stronger FP–salience coupling (slope B = +0.028 per SD of creativity, P = 0.022, 95% CI [0.004, 0.052]); in children the slope was slightly negative and non-significant (B = −0.014, P = 0.293). A parallel, trend-level pattern held for salience–VAN (adults B = +0.037, P = 0.062; children B = −0.025, P = 0.273). In adults, but not children, the neural machinery of creative engagement tracks measured creativity.
• Within-network group differences were concentrated in exactly the networks a co-creative dialog recruits. Talking with ChatGPT demands planning an utterance, monitoring both parties' responses, reorienting attention to unpredictable replies, and inhibiting off-task directions — the EF system. The adult-greater effects clustered in the frontoparietal control network, salience (which gates switching between internal and external attention), and dorsal attention — the same control/attention/modulation cluster the EEG ChatGPT study (Kosmyna et al. 2025) implicated behaviorally in adults.

What this means for practice

• "It looks easy" is not evidence it is easy for a 6-year-old. A five-minute ChatGPT conversation that feels effortless to an adult demands executive functions a first grader has not yet developed. The neural data suggest the task is not developmentally neutral: the same prompt produced a measurably different (and less integrated) engagement profile in children. Districts weighing chatbot exposure for early elementary students should treat "children use it and seem fine" as insufficient evidence.
• Age-of-use policy debates now have a neural data point. The authors explicitly position the findings alongside the debate on a minimum age for ChatGPT in schools, comparable to the policy attention given to screen exposure. The study does not set an age — but it reframes the question from "can children use it" to "when does the cognitive machinery catch up to the demand."
• Match the Scaffolding to the cognitive load, not the output quality. Children's matched creativity scores show they can produce; the under-integration of control networks shows the process differs. Interventions for young children using conversational AI should externalize the executive demands (structured prompts, planning scaffolds, adult co-engagement) rather than assuming the child supplies the planning, monitoring and inhibition internally.
• Creativity support and cognitive-control development are coupled in adults — an open question for children. In adults, creativity gains come with stronger FP–salience coupling; children show no such link yet. Longitudinal work is needed to find when the coupling emerges, which determines when AI-assisted creative work can be expected to "train" the same machinery it trains in adults.

Limitations

• The framing differed by group, and children's framing invited deference. Children were told ChatGPT was "the smartest computer in the world"; adults were simply told "ChatGPT." Children are known to defer heavily to perceived experts even when those experts are wrong, so the framing may have suppressed children's independent critical evaluation — the measured difference could reflect deference rather than (only) developmental capacity. The authors call for standardizing framing or manipulating perceived authority directly.
• No no-ChatGPT control condition in either group. The design compares children to adults under identical task demand, not ChatGPT use to no use. The findings describe developmental differences in ChatGPT dialog engagement; they do not isolate what ChatGPT itself does relative to human conversation or solo work.
• Small sample (n = 31) and motion confound. Fifteen children and sixteen adults is modest for imaging work, and despite the FD covariate the significant motion difference between groups (t(30) = 3.65, P < .001) means residual motion contributions cannot be fully excluded.
• Model version and collection window are not specified in the paper. The study reports "ChatGPT" without a model version or session dates. The findings describe the tool as tested in the study period, and replication on the current model generation is warranted before treating them as fixed properties of ChatGPT.
• Conversation content and instruction scope were narrow. A single five-minute open prompt (design a store) with mandatory verbal response, in a scanner, defines the boundary of what these results describe; generalization to classroom AI use — longer sessions, typed interaction, multi-turn instructional dialog — is not established.

Citation

Horowitz-Kraus, T., Kenett, Y. N., Link-Sourani, D., Ashqar, E., & Farah, R. (2026). Lower engagement of cognitive control in children on ChatGPT. Cerebral Cortex, 36(10), bhag167.