A 2026 meta-analysis of 51 studies found that incorporating ChatGPT into structured educational contexts produced large improvements in learning performance and moderate gains in higher-order thinking — and a separate brain imaging study at the University of Denver found that when kindergarteners interacted with ChatGPT alone, the brain region involved in understanding other people's minds was dramatically more activated than when children worked with a human.
Both findings are real, both published in peer-reviewed research, and they describe opposite ends of the same question: AI can accelerate learning and create unusual neural demands at the same time, and which one dominates depends almost entirely on how a child uses it and whether an adult is in the room.
This is the sixth piece in our ongoing series on AI and real life — following pieces on data privacy, AI health advice, wearable data, job displacement risk, and resume screening. This piece focuses on children specifically: what the research shows about how AI affects learning and brain development, where the evidence is strong, and where we genuinely don't know yet.
What the Research Shows: Benefits and Concerns Side by Side
| Research Area | What the Evidence Shows | Key Caveat | Evidence Quality |
|---|---|---|---|
| Learning performance (structured use) | Meta-analysis of 51 studies: large improvements in learning performance; moderate gains in higher-order thinking and learning perception | Results apply to structured educational settings — not free, unsupervised AI use | Strong — systematic review and meta-analysis |
| Academic self-efficacy | Structured AI use associated with enhanced academic self-efficacy and lower anxiety in young adult populations | Applies to adolescents and young adults more than young children; less data on under-10s | Moderate — emerging evidence |
| Brain activation (unsupervised use) | Kindergarteners alone with ChatGPT showed heightened activation in brain regions for understanding other minds; were more likely to treat AI as human-like | Study is under review, not yet fully peer-reviewed; small sample | Early — needs replication |
| Critical thinking (unsupervised use) | Studies on ChatGPT-assisted physics problem solving found unreflected acceptance of AI answers; reduced independent reasoning effort | Effect stronger when students use AI as answer-getter vs. thinking partner | Moderate — replicated across subjects |
| Children with disabilities | AI tools meaningfully improve accessibility and learning outcomes for children with diverse abilities | Implementation quality varies widely; teacher training is a bottleneck | Moderate to strong — consistent findings |
| Social-emotional development | Long-term effects still largely unknown; preliminary concern about chatbots disrupting normal social skill development if overused | No long-term longitudinal studies yet; current evidence is short-term | Weak — more research needed |
The Brain Imaging Finding That Deserves More Attention
The University of Denver study is the most striking recent finding in child-AI research, and it's worth understanding precisely. Professor Pilyoung Kim's research at the BAIC (Brain, Artificial Intelligence, and Child) center brought kindergarteners into the lab to co-create stories with ChatGPT under three conditions: children alone with the chatbot, children with a parent, and children collaborating with both a parent and ChatGPT together.
The findings were clear on two counts. First, children who interacted with ChatGPT alone showed heightened brain activity in the region associated with understanding other people's minds — the same neural machinery children use to figure out what another person is thinking and feeling. Second, those children were significantly more likely to treat the chatbot as human-like. Kim's interpretation: young children's brains work very hard to understand ChatGPT, in part because they're applying social cognition machinery that evolved for human-to-human interaction to a system that doesn't work the same way.
The implication for parents isn't "don't let children use AI." It's more specific: when a parent is in the room, the cognitive load is different and the interaction is healthier. When the parent was present, they helped bridge gaps in understanding — when ChatGPT didn't understand the child, the parent suggested another way to phrase it. The parental presence didn't diminish the child's engagement; it scaffolded it in a way that solo interaction didn't provide. That finding has immediate, practical implications for how AI use at home should be structured for young children.
The Learning Evidence: Strong Results With a Critical Qualifier
The meta-analysis of 51 studies reviewed by PMC/NLM — covering students from secondary school through university — found that structured AI use in educational contexts produced large improvements in learning performance and moderate gains in higher-order thinking. That's a meaningful finding. But the operative word is "structured," and it's doing a lot of work in that sentence.
The studies that showed learning gains were ones where AI was integrated into a curriculum with clear pedagogical intent — AI as a tutor that asks follow-up questions, AI as a tool to practice language output, AI as a system that adapts problem difficulty to a student's current level. The studies that showed reduced critical thinking and "unreflected acceptance" of AI answers were ones where students used AI as an answer-generator rather than a thinking partner — looking up the answer to a physics problem instead of working through the reasoning.
That distinction maps almost perfectly onto how children naturally want to use AI: the path of least resistance is to ask it for the answer, not to use it as a challenge. The PMC systematic review on AI use by children and adolescents noted that in young adult populations, structured use is associated with enhanced academic self-efficacy and lower anxiety. But it also flagged that unstructured use — particularly when children outsource the cognitive work rather than use AI to deepen it — shows the opposite pattern, with reduced independent reasoning effort over time. A child who consistently asks AI for the answer is practicing prompt submission, not problem solving.
The Age Question Matters More Than Most Coverage Suggests
The research doesn't treat "children" as a single category, and parents and educators shouldn't either. The effects of AI use vary significantly across developmental stages:
Early childhood (under 8): This is the age group with the least research and the most caution from neuroscientists. Early childhood is characterized by rapid synaptic formation and pruning — the brain is building the architecture for executive function, language, and social cognition. AI tools, if used in ways that replace rather than supplement human interaction, raise concerns about whether the right inputs are happening at the right time. Children in this age group also have the most difficulty distinguishing AI from a human interlocutor, as the Denver brain imaging study showed.
School age (8–12): This is where structured AI tutoring shows the strongest positive results. Children in this range have enough cognitive development to engage with AI as a learning tool rather than a social surrogate, and adaptive learning platforms — ones that adjust difficulty based on response patterns — have consistent evidence of supporting skill development in reading and math. The risk is homework shortcuts, and it's real: this is the age group most likely to ask AI to write their essay rather than use AI to improve one they drafted.
Adolescence (13–18): The evidence is most nuanced here. Structured AI use shows the largest positive effects on academic self-efficacy and learning outcomes. But this is also the developmental stage where social identity formation and peer relationship skills are most critical, and where excessive reliance on AI for social interaction — rather than human relationships — raises the most concern about long-term social development. Dr. Ying Xu from Harvard Graduate School of Education has specifically flagged the risk that chatbot interaction displaces the messier, more valuable work of navigating human social complexity.
Where Genuine Uncertainty Still Lives
It's worth being honest about how much we don't know, because this is a fast-moving area where confident claims often outrun the evidence.
The long-term developmental effects of growing up with AI as a daily tool are genuinely unknown. All of the studies showing learning benefits are short-term — weeks to a single academic year. We have no longitudinal data on what children who used AI extensively at age 8 look like academically and developmentally at 18. The brain region findings from the Denver study need replication. The question of whether children who outsource cognitive work to AI in school end up with weaker independent reasoning skills in adulthood hasn't been studied because those children don't exist as adults yet.
What neuroscientists do agree on: early childhood experiences have profound and long-lasting effects on brain architecture. The specific implications of AI interaction for that architecture are still being worked out, which is a reasonable basis for caution in how AI is introduced to very young children — not prohibition, but thoughtfulness about the type, structure, and supervision of AI use.
What This Actually Means for Parents
- Age under 8: be present. The Denver brain imaging research gives a specific, evidence-based reason for this — parental presence changes the nature of the interaction and reduces the cognitive confusion between AI and human interlocutors. AI isn't necessarily harmful for young children, but unsupervised AI interaction is the scenario with the least support and the most open questions.
- For school-age children: make AI a thinking tool, not an answer tool. The difference between "write me an essay about the Civil War" and "I've written this paragraph — what's unclear about my argument?" produces dramatically different cognitive outcomes. The first outsources the thinking; the second uses AI to deepen it. This distinction is worth making explicit with children, not just hoping they arrive at it themselves.
- For adolescents: monitor the social substitution risk. If an adolescent is increasingly using AI for conversation and emotional processing instead of working through those experiences with human peers and family, that's a pattern worth addressing — not because AI conversation is harmful in itself, but because the skill development that happens through human relationship navigation doesn't transfer from AI interaction.
- For all ages: talk about what AI is and isn't. Children who understand that AI doesn't have feelings, beliefs, or understanding — who have an accurate mental model of what they're interacting with — use it more effectively and are less likely to apply inappropriate social expectations to it. That understanding doesn't develop automatically; it needs to be taught.
Frequently Asked Questions
Is AI good or bad for children's learning?
It depends heavily on how it's used. A 2026 meta-analysis of 51 studies found large improvements in learning performance when ChatGPT was incorporated into structured educational contexts, while studies on unsupervised use found that children who used AI as an answer-getter showed reduced independent reasoning effort. The tool itself is neutral; the implementation determines the outcome.
Should young children be allowed to use AI chatbots?
Current research doesn't support a blanket prohibition, but it does support careful structuring. Brain imaging research from the University of Denver found that kindergarteners alone with ChatGPT showed heightened neural activation consistent with social confusion, while children whose parents were present showed healthier interaction patterns. Parental presence appears to be an important protective factor for very young children.
Will AI use hurt my child's critical thinking skills?
It can, if used to outsource thinking rather than deepen it. Studies on AI-assisted problem solving found "unreflected acceptance" of AI answers and reduced independent reasoning when students used AI to get answers rather than as a thinking tool. The risk is real but not inevitable — it depends on how the child is using AI and whether adults are helping establish the distinction.
What age is appropriate for children to start using AI tools?
There's no single evidence-based age threshold, but the research suggests more caution for children under 8 (where the brain is in rapid developmental phases and children have more difficulty distinguishing AI from human interlocutors) and more evidence of positive outcomes for structured AI use in children aged 8–12 and adolescents.
Can AI help children with learning disabilities?
Yes — this is one of the areas with the most consistent positive evidence. AI-powered personalized learning platforms, adaptive difficulty systems, and assistive tools have shown meaningful improvements in accessibility and learning outcomes for children with diverse abilities, including reading difficulties, attention challenges, and autism spectrum conditions.