What AI is doing to young brains

More than half of U.S. adolescents say they use AI chatbots to help with schoolwork, according to a 2026 Pew Research Center report. But that figure merely scratches the surface of how children and teenagers are interacting with AI—including everything from entertainment to companionship—and what that use might mean for their cognitive, social, and emotional development.

“Everything around youth and AI is essentially an open question right now,” said Anne Maheux, PhD, an assistant professor of psychology and neuroscience at the University of North Carolina at Chapel Hill who studies youth AI use. “There are certainly some positives, but experts tend to agree that the potential risks are more pertinent and urgent to understand.”

Psychologists like Maheux are working to understand not just whether children and teens are using AI, but if so how. They’re also launching research into pressing questions, including how AI tools can support learning rather than hinder it; what happens when children start to see chatbots as friends; and how to teach AI literacy skills so younger generations can learn when it’s safe or advisable to ask AI for help.

Lessons on AI, learning, and adolescent development

The following prompts (organized by topic) can support psychology educators in discussing how AI use affects learning, social development, and adolescent well-being in the classroom.

1. Learning, memory, and cognition
The article describes research showing that students who used an AI tutor saw significant grade improvements but performed no better than a control group once the tool was removed. Using what you know about encoding, retrieval, and the concept of desirable difficulty, explain why effortful learning may produce more durable knowledge than AI-assisted shortcuts. What does cognitive offloading theory suggest about the long-term risks of outsourcing learning tasks to AI? 
Reinforces: memory encoding and retrieval, desirable difficulty, cognitive offloading, and skill acquisition

Ultimately, psychologists are supporting broader efforts across society to integrate AI into learning environments with proper guardrails and to prepare today’s youth for a workforce and world infused with the technology.

“By the time today’s teens are adults, AI will be just as embedded in daily life as the internet—something we rely on so heavily we barely notice it’s there,” said psychologist Michael Robb, PhD, head of research at Common Sense Media, a nonprofit focused on how media and technology affect children and families. “The challenge will be figuring out where it fits without replacing the skills young people need to learn, grow, and function in society.”

According to a 2025 Anthropic report on AI productivity gains, general use large language models (LLMs)—such as ChatGPT, Claude, and Gemini—may enhance productivity on tasks such as writing and computer programming. But that comes at the expense of learning, likely because users engage less deeply with the material and do less active processing.

In one study of close to 1,000 high school math students in Turkey, grades improved 48% when students could use ChatGPT to help solve problems. But when ChatGPT was taken away, students performed 17% worse than those who never had access to the tool (Bastani, H., et al., PNAS, Vol. 122, No. 26, 2025opens in new window). To look at the mechanisms behind LLMs and learning, researchers from the Massachusetts Institute of Technology used electroencephalography to measure brain activity while participants wrote essays with or without ChatGPT’s help. They found that participants who used the LLM had weaker neural connectivity when completing the task. The findings have not yet been peer-reviewed (Kosmyna, N., et al., arXiv, 2025opens in new window).

[Related: How AI is reshaping human skills and thinking] 

“I believe this poses a significant risk for young learners: The more they grow up accessing quick solutions to problems, the less likely they will be to develop the requisite skills to solve those problems on their own,” said Shiri Melumad, PhD, an associate professor of marketing at the Wharton School of the University of Pennsylvania.

That raises a broader question: If AI can perform certain tasks with near-perfect accuracy, do young people still need to learn those skills? Learning science experts say yes, emphasizing that basic skills are needed to acquire higher-order ones. For example, students still learn how to do arithmetic by hand despite easy access to calculators because it helps them build the foundation for more advanced mathematical reasoning. Along the same lines, learning to code still drills useful cognitive skills such as logical structuring, breaking problems into steps, and computational thinking—even if AI increasingly takes over coding tasks.

“That’s the argument for why we shouldn’t offload these critical patterns of thought. If students don’t already have these habits of mind, they can’t build on top of them,” said Zachary A. Pardos, PhD, an associate professor of education at the University of California, Berkeley, who studies adaptive learning and AI.

Sometimes called a “productive struggle,” learning can be a laborious process where students must wrestle with challenging material to improve their performance. The idea is rooted in psychologist Robert Bjork, PhD’s concept of desirable difficulty—a learning task that demands considerable effort but offers long-term benefits.

“If your goal is learning, then efficiency and ease should take a back seat,” said Brooke Macnamara, PhD, an associate professor of psychological sciences at Purdue University who studies how relying on AI may weaken skills over time through a process called cognitive offloading, where we use outside tools to reduce the mental effort required for a task.

When a student gets instant gratification from AI’s quick answers, they also miss the chance to develop skills like patience, perseverance, and emotion regulation, Robb said.

In a survey, Maheux asked young adults how often they outsource entire tasks to AI. Those who did so more frequently reported more anxiety (PsyArXiv, 2026opens in new window). In unpublished data, they also showed declines in distress tolerance over time.

“Distress is not always bad. Learning to regulate emotions is a critically important part of adolescence, but the cognitive offloading that happens with AI use may be undermining the development of that skill,” Maheux said.

LLMs are so adaptive and open-ended that the user experience can vary substantially from one person to the next. For that reason, researchers stress that what matters is not whether students use AI but how and why they use it.

Even standard LLMs such as ChatGPT can have a place in education when they are used to support certain cognitive tasks, Pardos said. For example, students can use these tools to offload nonessential tasks such as formatting, generating boilerplate text, or searching for information. They can also serve as brainstorming partners, create sample math problems along with a step-by-step breakdown of solutions, or provide critical feedback on a draft written by the user.

“Those uses are very different from completely automating and outsourcing a learning task,” Maheux said.

Psychological science can help guide classroom decisions about AI by identifying which learning experiences are essential for building lasting knowledge and skills—and which tasks AI can safely support.

Growing evidence also suggests that generative AI with guardrails may serve as an effective tutor. In the study of high school math students in Turkey, researchers tested a custom LLM with prompts designed to support learning. Students who used the tutor saw even greater gains—a 127% improvement in grades, compared with 48% for those using ChatGPT. When the tutor was removed, these students performed just as well as a control group, avoiding the learning losses seen in the ChatGPT group. On the other hand, students who used the custom tutor appeared to overestimate their learning and test performance, suggesting that even AI with guardrails may create an illusion of mastery (Bastani, H., et al., PNAS, Vol. 122, No. 26, 2025opens in new window).

Intelligent tutoring systems, such as McGraw Hill’s ALEKS and Carnegie Learning’s MATHia, already use AI to analyze student performance and adjust the pace, difficulty, and level of support accordingly. Integrating LLM-driven chatbot support could further enhance these systems by breaking problems into smaller steps, asking guiding questions, and explaining mistakes in real time, said Pardos, who led the creation of an open-source intelligent tutoring system called OATutoropens in new window.

Generative AI tools could also help integrate special education with general education by making it easier for students with different needs to learn in the same classroom, said Tammy Kwan, EdD, an entrepreneur in residence at the Princeton University AI Lab. For example, the AI tutoring tool known as Read Sidekick, a finalist in the Stanford Accelerator for Learning’s 2026 Create+AI Challenge, helps people with reading difficulties access and understand complex text. The free browser extension gives line-by-line interpretations and reading instruction and will soon add an English-to-American Sign Language translator with an animated signing avatar.

“This is a really exciting direction for AI and augmented learning,” Kwan said. “If we’re able to do it right, generative AI can help make the education system more inclusive.”

Outside the classroom, children and teens are also using what’s known as “social AI” for entertainment, companionship, and advice. This includes activities like listening to a story, asking for advice on a tricky interaction, or practicing face-to-face conversations. These interactions can happen on specialized apps such as Replika or Nomi or general LLMs such as ChatGPT.

Surveys from Common Sense Media show that 72% of children have used social AI (Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions, Common Sense Media, 2025opens in new window). About a third of those children say they have discussed serious problems with AI instead of talking to another human about them.

“These uses may make parents uncomfortable because they can replace roles that would previously have been filled by another person,” Robb said. “We’re very interested in understanding what downstream effects that might have.”

[Related: Artificial intelligence and adolescent well-being]

In some cases, social AI may be beneficial—for instance, when a child has limited access to peers or support networks (for example, because of a remote location) or when neurodivergent children want to practice offline interactions. Experts caution that these tools should supplement—rather than replace—real-world relationships. Significant risks also exist, Maheux said, and most of the potential benefits remain speculative. In addition to displacing human relationships, use of social AI may interfere with the development of key social skills such as talking synchronously and interpreting body language.

Another major concern: AI chatbots tend to be warm, positive, and sycophantic, so interactions with them generally lack friction, Robb said. Children who use social AI may therefore miss important opportunities to take another person’s perspective, deal with disagreement, or resolve a conflict.

“Because they are not sentient creatures, AIs have no needs of their own, so the child has no opportunity to practice prosocial skills, empathy, or conflict negotiation,” Maheux said. “For those reasons, it’s possible that AI could lead to more dysfunctional beliefs about how human relationships should be.”

Though the results are not yet published, her research suggests that youth who hold more dysfunctional relationship beliefs—such as the idea that people should be able to read each other’s minds—may develop more interest in companionship-style relationships with AI over time.

Those trends are particularly concerning given that adolescence is a sensitive window for social-emotional development, sometimes called the “age of opportunity” because the teen brain is uniquely plastic for building social and emotional skills.

“If these skills are outsourced or undermined, young people could face challenges in these areas later in life,” Maheux said.

While some AI tools require age verification, many popular LLMs allow anyone to log in and start chatting. Experts say adding better safeguards could help protect young people until research reveals what sort of use is problematic. For example, a toggle could allow parents, teachers, students, or users under a certain age to switch to a structured experience that functions more like a tutor than an open-ended conversation partner, said Mutlu Cukurova, PhD, a professor of learning and artificial intelligence at University College London who studies human-AI interaction.

“My hope is that as we increasingly invest in research, we take a more precautionary approach with children’s use of AI until we have evidence that it is safe for them to develop in that kind of environment,” said Nita Farahany, PhD, JD, a professor of law and philosophy at Duke University who studies the legal, ethical, and social aspects of emerging technologies.

Parents also have a role to play in keeping children safe. Some may feel they don’t fully understand AI and therefore don’t know how to discuss it with children. But opening up channels of communication—asking whether children are using these tools and how—can help parents spot problems early and guide healthier boundaries.

“We know from digital well-being research that parent mediation matters,” Robb said. “Kids who have active conversations with parents about how they’re using technology, including AI, are better positioned to develop healthy habits around it.”

Schools need to provide clear, well-developed policies on how AI can and cannot be used for schoolwork so students are not left to fill in the blanks, he added.

Another priority for schools should be teaching students to use AI responsibly. AI literacy can help kids learn the risks and benefits and start to choose for themselves what constitutes safe AI use. That will require changes to the educational infrastructure and professional development so teachers feel equipped to guide students, Kwan said. Common Sense Media offers free AI literacy lesson plansopens in new window and other resources for educatorsopens in new window.

“Teaching students when not to use AI—and requiring reflection on their choices—may help mitigate the risks around healthy development,” said Nicole Barnes, PhD, APA’s executive lead psychologist for education.

Ongoing research seeks to further understand how AI can support, rather than hinder, learning and development. Cukurova is exploring an area known as AI-human complementarity, where new capabilities can support interactions that are more synergistic. For example, AI models with better “theory of mind” abilities—or the awareness of what someone is thinking or feeling—may be able to challenge users rather than pandering to them.

Maheux hopes researchers can learn from the social media literature and take a more effective approach to studying youth AI use. Early social media research focused heavily on screen time but later showed that how young people use technology matters much more than how often they use it. For AI, that means getting precise about what kids are actually doing. Maheux’s group is currently building a taxonomy of ways children and teens are using AI. She said longitudinal studies will be key for understanding the technology’s effects.

“We need to know what comes first—kids’ vulnerabilities or something about the AI use itself,” she said. “If we’re not looking at both short- and long-term time scales, we may miss the bigger picture of how these technologies influence development.”