脑成像研究比较儿童与成人和 ChatGPT 对话时的脑网络

比较的是同一项对话任务

这项研究关注的不是 ChatGPT 是否让儿童变得更有创造力，而是儿童和成人在与聊天机器人共同构思时，脑内与控制、注意有关的网络如何协同。研究团队让15名平均约6.4岁的儿童和16名平均约38岁的成人，在功能性磁共振扫描中用口头对话规划一家商店的开场。对话持续五分钟，参与者可以自行决定商店的类型、地点和服务对象。

扫描前，参与者还完成了词语联想创造力测试；研究者也把对话转录，估算其中词义的多样程度。两种测量回答的是不同问题：测试反映扫描外的一项基线创造力表现，对话指标反映这次交流中语言内容的语义跨度，而脑成像记录的是网络活动的同步。不能把其中任一项直接等同于整体创造力或实际学习成效。

表现相近，部分网络连接不同

在扫描外的测试中，儿童和成人相对于各自年龄组的平均百分位接近，差异没有达到统计显著。实际对话中，成人的语义整合分数略高于儿童，分别为0.82和0.81；报道将这一小幅差异称为统计上可靠。脑成像分析则发现，成人的额顶控制网络内部同步更强；经多重比较调整后，跨网络差异中较稳健的一项，是额顶控制网络与背侧注意网络的连接在成人中更强。

这些结果呈现的是年龄组在这项任务中的差别，而不是儿童缺乏创造力。研究还发现，成人的基线创造力与对话时额顶控制网络和显著性网络之间的连接呈正向关联；儿童组没有显著关联。不过作者指出，这类创造力与连接的分析属于探索性分析，样本较小，可能统计效力不足，不能据此断定两组的创造力机制已经被完整解释。

不能归因于 ChatGPT 的独特影响

最重要的解释边界是研究没有让参与者在扫描中独自头脑风暴，也没有安排他们与真人伙伴交谈。因此，儿童在部分网络连接上的较低水平，可能反映一般的发展差异或任务难度，而不是 ChatGPT 特有的作用；研究不能证明聊天机器人降低了儿童的认知控制，更不能证明它损害脑部。

实验说明也并不完全一致：儿童被告知将与“世界上最聪明的电脑”交谈，成人则只被告知将与 ChatGPT 交谈。研究者认为，对 AI 权威性的不同提示可能影响儿童是否质疑或评估回应。此外，样本人数有限、两组性别构成不均衡，日常词汇量、说话速度、对话轮数和此前使用 ChatGPT 的经验也未被控制。报道所引研究团队因此提出，后续需要扩大样本、比较 AI 对话与真人对话或无 AI 构思，并进行长期追踪。

研究者建议成人陪伴儿童使用生成式 AI，但这属于对结果的实践解读，不是本研究直接检验过的干预效果。现有发现提示的是：在这项特定对话任务里，儿童与成人表现出不同的脑网络协同模式；它尚不能回答何时适合开始使用 AI，或成人陪伴是否会改善交流表现。

A new brain imaging study suggests that young children show less communication within and between the brain networks responsible for self-regulation and focused attention when brainstorming with an artificial intelligence chatbot than adults do. Although first graders and adults scored similarly, relative to their age groups, on a standard test of creative thinking outside the brain scanner, only adults showed a link between higher creativity and stronger coordination among these mental control systems during the conversation. The findings, published in Cerebral Cortex, suggest that young children may not yet have the mature brain wiring needed to actively direct and evaluate a collaborative conversation with generative artificial intelligence.

Generative artificial intelligence programs, such as ChatGPT, are powered by large language models, which are computer systems trained on massive amounts of text to predict how words fit together, allowing them to produce human-like responses and brainstorm ideas in real time. As these tools enter homes and elementary classrooms, children are increasingly using them for storytelling and collaborative problem-solving, a process known as co-creativity. Keeping up with a fast-talking, unpredictable digital partner requires cognitive control, which is the mind’s ability to plan ahead, stay focused on a goal, remember what was already said, and filter out unhelpful distractions.

Generating original and useful ideas relies heavily on how different systems in the brain communicate with one another. For instance, a 2018 neuroimaging study of adults (covered by PsyPost at the time) indicated that people who come up with more original ideas tend to have stronger synchronization among three major brain networks: one involved in spontaneous daydreaming and memory, another that flags important new thoughts, and a third that handles cognitive control and goal-directed planning. Instead of operating separately, these systems team up so a person can both brainstorm freely and judge whether an idea actually makes sense.

In young children, however, the brain networks that manage cognitive control and focused attention are still under construction, only reaching full maturity in a person’s mid-twenties. A 2019 neuroimaging study of young children, conducted by the same research group behind the new study, found that across children ranging from 17 months to nine years old, older children tended to show stronger connections within and between their cognitive control networks while listening to stories. In addition, a neuroimaging study covered by PsyPost in 2023 indicated that the way cognitive control networks relate to creative thinking shifts across development, with stronger connectivity in these control regions linked to higher creativity in adolescents but lower creativity in adults.

First grade, around ages six and seven, is a sensitive developmental window when children are first learning to evaluate text and speech on their own, while still tending to trust perceived experts and technology without much skepticism. The research was led by Tzipi Horowitz-Kraus, an associate professor with the Educational Neuroimaging Group in the Faculty of Education in Science and Technology and the Faculty of Biomedical Engineering at the Technion – Israel Institute of Technology, who is also affiliated with the Kennedy Krieger Institute and Johns Hopkins University School of Medicine.

She and her Technion colleagues Yoed Kenett, Dafna Link-Sourani, Emeel Ashqar, and Rola Farah examined whether first graders engage their cognitive control and attention networks to the same degree as adults when brainstorming with ChatGPT, and how that brain activity relates to their underlying creative ability. Part of the motivation came from a 2025 preprint, which has not been peer reviewed, in which adults who wrote essays with ChatGPT showed less coordinated brain activity than those who used a search engine or no tools at all.

“Generative AI is entering children’s lives, and classrooms, much faster than the evidence about how young brains engage with it,” Horowitz-Kraus told PsyPost. “Studies in adults have suggested that people recruit fewer cognitive resources when they work with ChatGPT, but adults have mature cognitive control systems, and these systems keep developing into the mid-twenties. We focused on first graders (ages 6–7) because this is when formal schooling begins, when executive functions are developing rapidly, and when children tend to place a great deal of trust in sources they perceive as experts. We wanted to know what the brain of a young child is actually doing during a real conversation with ChatGPT, compared with an adult given exactly the same task.”

To investigate this, the researchers recruited 15 first-grade children with an average age of about 6.4 years and 16 adults with an average age of about 38 years. All participants came from average to above-average socioeconomic backgrounds, had no history of neurological or psychiatric conditions, and completed standard tests to confirm typical verbal and nonverbal intelligence. Before entering the brain scanner, each participant took a behavioral creativity test called the Divergent Association Task, which asks a person to name 10 words that are as unrelated in meaning as possible.

Inside a functional magnetic resonance imaging (fMRI) scanner, which measures brain activity by tracking changes in blood flow, participants engaged in a five-minute spoken conversation with ChatGPT. Using a specialized microphone and noise-canceling audio system built for the scanner, both children and adults were given the same open-ended goal: they had to plan the opening of any kind of store they wanted, anywhere and for anyone, by consulting out loud with ChatGPT. Immediately before the scan began, children were told they would be speaking to “the smartest computer in the world,” also known as ChatGPT, whereas adults were simply told they would be speaking to ChatGPT.

“I would also like to note that holding a live spoken conversation with ChatGPT inside an MRI scanner, with six-year-olds, required dedicated infrastructure developed at the Technion Human MRI Research Center,” Horowitz-Kraus noted. “I am very grateful to the children, parents and students who took part.”

The researchers recorded and transcribed every conversation to measure how original and varied the spoken exchanges were. They used a computer model to calculate a “divergent semantic integration” score, which measures how far apart the meanings of all the words in a conversation are from one another; for example, the words “cat” and “democracy” sit much farther apart in meaning than “cat” and “dog.” On the brain scans, the team measured functional connectivity, which reflects how closely different brain regions synchronize their activity over time, across seven major brain networks after statistically controlling for head movement inside the scanner.

These seven systems included the frontoparietal and cingulo-opercular networks, which manage cognitive control and task goals; the dorsal and ventral attention networks, which handle voluntary focus and automatic reactions to new information; the salience network, which acts as a switchboard by detecting important ideas; the default mode network, which supports spontaneous imagination; and the memory retrieval network.

On the word-association creativity test taken outside the scanner, children and adults scored at comparable levels relative to others their age, with average percentile scores of 28.61 and 26.16, respectively, a difference that was not statistically significant. During the actual five-minute conversation with ChatGPT, however, adults produced slightly more diverse and wide-ranging spoken narratives than the children did, scoring an average of 0.82 on the semantic integration measure compared to 0.81 for the first graders, a small difference that was nonetheless statistically reliable.

When the researchers examined the brain scans, they found pronounced differences in how the two age groups engaged their mental control and attention systems while talking to the chatbot. Within individual brain networks, adults showed substantially stronger internal synchronization in the frontoparietal control network, where adults averaged a connectivity score of 0.201 compared to 0.113 in children. Adults also had higher internal connectivity in the salience network (0.289 versus 0.198 in children) and the dorsal attention network (0.211 versus 0.162 in children). Only the frontoparietal difference held up after the researchers adjusted for the number of comparisons they ran.

In contrast, internal connectivity within the default mode network, which handles spontaneous imagination, did not differ between the two groups, averaging 0.187 in adults and 0.181 in children. Communication across different brain networks was also more integrated in adults than in first graders during the AI conversation. After adjusting for multiple statistical comparisons, the most robust difference appeared in how the frontoparietal control network linked up with the dorsal attention network (averaging 0.126 in adults versus 0.039 in children).

Before correcting for multiple comparisons, adults additionally showed stronger cross-network ties linking memory retrieval regions to cognitive control and attention systems, as well as stronger coupling between the default mode network and memory retrieval (0.150 in adults versus 0.111 in children). The researchers then tested whether a person’s baseline creativity score outside the scanner predicted how strongly their brain networks communicated while using ChatGPT.

They found that in adults, higher creativity scores were positively linked to stronger communication between the frontoparietal control network and the salience network during the AI dialog. Among the first graders, however, baseline creativity showed no statistically significant relationship with the strength of these brain network connections. That adult pattern differs from the 2023 study described earlier, which tied stronger connectivity within the executive control network to lower creativity in adults, though that study looked at connections within a single network rather than links between networks.

“Two things” surprised the team, Horowitz-Kraus explained. “First, children and adults looked similar in creative ability on the standard test, yet the neural route they took during the conversation was quite different; the difference lay in how creativity is implemented in the brain, not in baseline capacity. Second, the default mode network, which is central to most models of creativity, did not differ between the groups. One possible explanation is that conversing with ChatGPT in a scanner was novel for everyone, but that needs to be tested directly.”

Taken together, the results suggest that even when young children have age-appropriate creative potential, they engage their control, attention, and memory systems in a less integrated way than adults do during a conversation with generative AI. The authors propose that adults use their salience and frontoparietal networks to filter ChatGPT’s rapid responses and steer the brainstorming toward their goal, a role these networks may not yet play in six- and seven-year-olds.

“Talking to a chatbot looks effortless, but it draws on planning, monitoring, attention and memory systems that are still under construction in a six-year-old, which is why adult guidance matters when young children use these tools,” Horowitz-Kraus said.

“The practical message is that young children should not be left to use generative AI on their own,” she added. “For parents and teachers, this means sitting with the child, asking questions about the answers, and treating the AI as a shared activity instead of an independent tutor. For educators, pediatricians and policy makers, it means that the age at which children start using these tools deserves the same careful attention we already give to screen exposure. This study cannot set a specific ‘safe age’ since it focused on two very distinct age groups, but it is a clear signal that tools designed around adult abilities should not simply be handed to young children.”

There are a few things to keep in mind when interpreting these findings. Because the study did not include a comparison task where participants brainstormed on their own or talked with a human partner inside the scanner, the design cannot separate general age differences in brain maturity from effects caused specifically by ChatGPT. In other words, the lower network connectivity seen in children may reflect how six-year-old brains handle this kind of complex task in general, rather than showing that ChatGPT actively suppresses or harms their brain activity.

“This study does not show that ChatGPT reduces cognitive control in children or harms their brains,” Horowitz-Kraus emphasized. “We compared children with adults doing the same ChatGPT task; we did not compare children with and without ChatGPT, or ChatGPT with a human conversation partner. So we cannot yet separate what is specific to AI from general developmental differences in how the brain handles this kind of conversation.”

The way the chatbot was introduced may have also influenced how the children behaved. Telling first graders that they were about to talk to “the smartest computer in the world” may have encouraged them to passively accept the machine’s suggestions rather than questioning or evaluating its ideas, which could naturally lower their use of cognitive control networks. Future research will need to give children and adults the exact same instructions, or intentionally test how calling an AI an “expert” changes how critically kids think.

In addition, the sample size of 15 children and 16 adults was relatively small, and the groups were unevenly split by sex, with 11 boys among the 15 children and four men among the 16 adults. The study had enough participants to detect large differences between age groups, but the analyses linking creativity to brain connectivity were exploratory and likely underpowered.

The researchers also did not control for differences in everyday vocabulary size, how fast participants spoke, how many turns they took in the conversation, or how much prior experience they had with ChatGPT before the experiment (and adults likely had far more experience with the tool). Tracking children over several years in larger studies will help educators and pediatricians determine the developmental stage at which kids build the mental control needed to actively collaborate with artificial intelligence tools.

“The next steps are larger samples, a direct comparison of talking with ChatGPT versus talking with a human partner or working without AI, and longitudinal follow-up to identify when in development the control networks begin to engage in an adult-like way and to relate to creativity,” Horowitz-Kraus said. “We also plan to measure children’s familiarity with AI and to test how the perceived authority of the AI shapes their engagement. The long-term goal is to give educators, pediatricians and policy makers evidence for developmentally appropriate use of generative AI, in the same way that guidance exists for screen exposure.”

“Technology can support learning, but it cannot lead it; children learn best alongside an adult who guides and challenges them,” she added.

The study, “Lower engagement of cognitive control in children on ChatGPT,” was authored by Tzipi Horowitz-Kraus, Yoed N. Kenett, Dafna Link-Sourani, Emeel Ashqar, and Rola Farah.