Scientists from Israel's Technion have found that in children aged 6-7, during a conversation with ChatGPT, the brain areas responsible for self-control and attention interact more weakly than in adults. The work was published in the journal Cerebral Cortex. It involved 15 first-graders (mean age 6.4 years) and 16 adults (mean age 38 years).
Participants performed the same task: over five minutes, aloud, they used the chatbot to come up with a concept for opening any store. Before starting, the children were told they would be talking to "the smartest computer in the world," while the adults were told simply that they would be talking to ChatGPT. The conversations were recorded and analyzed. Brain activity was measured using functional magnetic resonance imaging (fMRI), a method that records changes in blood flow in different areas.
Before scanning, all participants took a divergent thinking test: they had to name 10 words as far apart in meaning as possible. Compared with their peers, the children and adults showed comparable results (mean percentiles 28.61 and 26.16 respectively, the difference statistically insignificant). However, in a live conversation with the bot, adults produced slightly more varied utterances: 0.82 versus 0.81 on the metric of semantic spread of words - the difference is small but statistically reliable.
The main differences were found in the connections between neural networks. The internal coherence of the frontoparietal network (responsible for planning and control) was higher in adults: 0.201 versus 0.113 in children. After correction for multiple comparisons, this was the difference that remained significant. Adults also had a stronger connection between the frontoparietal network and the dorsal attention network (0.126 versus 0.039). And the default mode network, associated with imagination, did not differ between the groups in internal connectivity (0.187 in adults and 0.181 in children).
In addition, in adults, higher creativity scores were associated with strengthened connectivity between the frontoparietal network and the salience network (which helps notice important ideas). No such connection was found in children. This is a correlation - a statistical relationship that by itself does not prove causation.
The authors emphasize: the study does not show that ChatGPT reduces control in children or harms the brain. It did not include a comparison with a conversation without the bot or with a person, so it is impossible to separate features specifically related to AI from general age differences. The sample is small, and the groups are uneven by sex. In addition, children and adults were given different instructions: the phrase about "the smartest computer" may have encouraged children to accept the bot's answers more passively. The scientists plan to increase the sample and determine when control systems begin to work in an adult-like way.
"Talking with a chatbot seems simple, but it relies on planning, attention and memory, which are still developing in a six-year-old, so adult accompaniment is important," noted the head of the work, Tzipi Horowitz-Kraus. She warned against leaving young children alone with generative AI.