‘Cognitive stunting’ and the potential threat generative AI poses for children’s brains

Los Angeles Unified School District recently announced it will bar student use of generative artificial intelligence on school devices for the 2026-27 academic year, joining New York City Public Schools in limiting children’s access to a technology no one yet knows how to safely manage.

The nation’s two largest school districts made the move at a pivotal moment in the technology’s evolution. Last week, a former engineer for OpenAI and Anthropic warned that AI’s chief architects “earnestly believe that it could kill us all by the end of the decade.”

It’s been less than four years since the public debut of ChatGPT, the first widely adopted generative AI platform. Given the relative novelty of publicly-accessible generative AI platforms, the total body of research studies investigating their effect on children’s cognitive development is fairly small.

But the available data paint a worrisome picture of thwarted learning.

Without careful guardrails in place, children tend to use the platforms as shortcuts to answers, bypassing the messy, difficult cognitive processes that are essential for learning.

Researchers watching young children’s brain activity as they interacted with ChatGPT saw less engagement in regions associated with attention and information processing than they did in adults doing the same task. High school students permitted to use ChatGPT as a study aid soared through practice questions, but performed worse on tests than peers without AI once the tool was taken away.

One researcher investigating AI’s effects on children has expressed concerns of “cognitive stunting” — critical thinking skills prohibited from developing, just as physical growth can be stymied by malnutrition.

It’s appropriate to be cautious, researchers said.

“I believe this moratorium is the right decision with the current evidence we have,” said Anne Maheux, an assistant professor of psychology at the University of North Carolina at Chapel Hill who studies youth and AI.

“These tools are still being built, evolving in technological capacity before our eyes, and designers are responding to current profit incentives in ways that may not always be aligned with promoting positive youth development,” she said. “Evidence in support of a particular tool’s benefits should be extremely clear before these tools are rolled out at scale for children and teens.”

Adults’ performance on a given task tends to diminish the more they outsource that job to technology, a phenomenon known as skill erosion. It’s a trade-off that has accompanied virtually every new technological development. (According to Plato, Socrates famously groused that students’ preference for books over memorized text would “implant forgetfulness in their souls.”)

But generative AI has vastly broadened the range of tasks that humans can successfully outsource. That’s a boon to professionals looking for virtual assistance with work they already know how to do. For students who haven’t yet mastered a skill, it can be an obstacle to learning.

“When young people systematically use AI to complete schoolwork or other thinking tasks, they are not offloading skills they already possess. They are shortcutting the process of developing those skills in the first place,” wrote Rebecca Winthrop, a senior fellow and director of the Center for Universal Education at the Brookings Institution.

Winthrop coined the term “cognitive stunting” to describe this effect on children.

The limited research available offers some indications that children and adults process AI interactions differently.

A team from Technion-Israel Institute of Technology carried out an experiment in which 15 6- and 7-year-olds and 16 adults first sat for a short test to assess their baseline creativity skills. Participants then did an exercise designing a store of their choice using ChatGPT, while fMRI scanners recorded their brain activity.

For adults, their creative output with ChatGPT was generally on par with the creativity level they’d shown on the test.

Children who scored high on the creativity test, however, produced imaginary stores no more creative than children who scored lower. The fMRI also detected less synchronized blood flow in regions of the children’s brains associated with attention and information processing than they did in the adults’.

The study was released on a preprint server and is still undergoing the process of peer review. Its scope is limited, but to co-author Tzipi Horowitz-Kraus, an associate professor and head of Technion’s Educational Neuroimaging Group, said the findings suggest that “six years old [is] probably too early to use generative AI tools,” she said. “When is a good time? We don’t know. More research is needed.”

A paper published last year offers more insight into AI’s effect on older students. One group of high school students in Turkey were allowed to use a generative AI platform similar to ChatGPT when studying for a math test. A second group studied without AI access. A third was allowed to study with a modified AI platform designed by their teachers to offer hints without ever providing answers directly.

The students with AI access did great on practice questions, outperforming those without it by 48% (those using the general platform) and 127% (those using the tutor).

When it came time to take an offline test, the students who used the general platform, which was capable of giving them the answers in practice, scored 17% lower than those with no AI. Those who used the AI tutor did as well — but no better — than AI-free kids.

When left alone with a tool capable of doing practice questions for them, the students used AI “as an answer machine,” said co-author Alp Sungu, an assistant professor of operations, information and decisions at the Wharton School of the University of Pennsylvania.

Sungu is a supporter of cellphone bans in schools, and leans toward banning social media for children as well, he said. He has yet to see a paper demonstrating that AI access improves learning outcomes in children. At the same time, he is ambivalent about the efficacy of outright bans on school devices.

It isn’t that he thinks that children should have unlimited access to AI. Rather, a school device is the only place educators can directly control the material students are exposed to. A moratorium can’t prevent students from using ChatGPT and its ilk on home devices, but it does reduce opportunities to show them how to use it wisely.

“The goal should be about making students use it for productive purposes, and that can only happen [by] interacting with students,” Sungu said.

LAUSD officials will spend the coming months working on policy recommendations and guardrails to present to the school board at the end of the academic year, a spokesperson said.

Given the speed of AI’s evolution, it’s possible students will be grappling with even more powerful platforms before educators and policymakers have fully reckoned with the ones they already have, said Matthew Kopec, director of Harvard University’s Embedded EthiCS program.

“The tools are evolving so quickly that students are effectively using very different tools by the time the possible negative effects of the old tools start appearing,” Kopec said. “I think we have many very good reasons to heavily regulate how generative AI tools are allowed to interact with our children.”