Why do children's brains show weaker control networks during AI dialogue?

An fMRI study scanned 15 six-to-seven-year-olds and 16 adults during a naturalistic 5-minute co-creative dialogue with ChatGPT, using the same prompt for both groups, and measured functional connectivity (Fisher's z) within and between seven large-scale brain networks — cingulo-opercular, default mode (DMN), memory retrieval, frontoparietal (FP), salience, ventral attention, and dorsal attention (DAN). "Creativity scores, assessed outside of the scanner, did not differ between groups," but "adults showed stronger within-network connectivity in control/attention systems, most robustly in FP, with additional effects in salience and DAN," plus "greater cross-network integration, particularly FP-salience and FP-DAN couplings." The paper's own framing: "children exhibit a less coherent and integrated control–attention–memory architecture than adults" while using the tool.

The mechanism the authors invoke is the standard creativity-network account, applied to this task. The DMN supports memory retrieval and "remote associative thinking," generating candidate ideas and evaluating them against stored knowledge; the frontoparietal executive-function (EF) system provides top-down regulation — inhibiting "dominant but less original responses" so more novel ones surface; and the salience network "orchestrates interactions between the DMN and the EF system, dynamically transitioning between generative and evaluative modes." Prior connectome work ties stronger DMN-EF-salience coupling, and more frequent transitions between DMN and EF, to higher creative ability in general (not AI-specific) tasks. Applied here: adults showed the FP-salience and FP-dorsal-attention coupling this account predicts for coordinated, goal-directed idea generation during the AI dialogue; children's networks were measurably less integrated on the same task. A further result sharpens this: "in adults, but not in children, higher creativity was associated with stronger connectivity within cognitive control networks" — the brain-behavior link the theory predicts holds for adults but did not show up in the children's data.

This sits alongside Does AI assistance weaken our brain's ability to think independently?, which found neural coupling scales down with AI assistance across a four-month longitudinal writing study in adults. That study varied the amount of AI support and watched connectivity drop as support rose; this one holds AI support constant (same brief ChatGPT dialogue) and varies developmental stage, finding children's networks less integrated than adults' from the outset. The two point at different sources of "weaker" neural engagement during AI use — accumulated reliance versus a still-maturing control architecture — that could compound in a child who grows up co-creating with AI tools, though this excerpt does not test that combination.

What the study does not establish: it has no non-AI condition, so it cannot say whether this child-adult connectivity gap is specific to AI-assisted co-creation or simply reflects normal frontoparietal and attention-network maturation present in any goal-directed creative task at this age. The samples are small (15 and 16) and the task is a single 5-minute session, so the finding describes this sample's connectivity pattern rather than a general claim about "children and ChatGPT." The authors' own gloss — that matched creativity scores alongside divergent neural profiles means "developmental differences in the implementation of creative cognition may be more pronounced than baseline capacity" — is the right-sized conclusion: a difference in neural architecture, not yet evidence of a difference in what children can actually produce.