AI in Education: What the Research Says

Broadly defined, “AI” refers to technology that allows computers to mimic aspects of human thinking, problem solving, and learning, says Mark Warschauer, PhD, Distinguished Professor of Education at the University of California, Irvine. AI is increasingly used in educational settings, often in different forms that work together:

• • Generative AI – (e.g., ChatGPT) can create text, images, lesson materials, and even computer code in response to user prompts.
  • Non-Generative AI – 
    • Adaptive: adjusts to a student’s progress and performance to help personalize learning. These systems are used in educational platforms such as DreamBox, i-Ready, and ALEKS.
    • Assessment & Evaluation:  automated grading
    • Intelligent Tutoring systems: teaching content and providing feedback
    • Profiling & Prediction: making admission decisions, predicting drop-outs, etc

(Mark Warschauer, Zawacki-Richter et al., 2019)

Research shows that students lead other groups in AI adoption, and ChatGPT has stated that students ages 13-24 in the United States are its largest single demographic group, says Amanda Bickerstaff, MS Ed, Co-Founder and CEO, AI for Education.

Upwards or 90% of high school students say that they use some sort of generative AI for their schoolwork, says Adam Dubé, PhD, Associate Professor of Learning Sciences at McGill University. Dubé’s research also indicates that over half of Canadian families with children in grades kindergarten through grade four are using AI-enabled math apps every week.

Current research indicates only about 10% of students use AI for cheating consistently, says Dubé. This is consistent with data from the pre-AI era that a similar percentage consistently cheat on exams or have someone else do their homework.

However, large numbers of students are using AI in ways that are in a bit of a gray zone. Dubé’s research shows:

• • 80% of students say they use it to explain ideas to them because they’re confused and they want an extra explanation. 
  • 70% say that they use it to generate ideas
  • 60% of them say that they use it to summarize a text instead of reading it. 
  • About 45%  are using it to edit portions of what they write. 

While none of these uses would be considered “cheating,” notes Dubé, it’s worth asking if getting this help is allowing the student to develop fundamental skills they need to learn instead of having the AI system do it for them.

The short answer – it is nearly impossible to know. The omnipresence of AI in every app and every computer – and the ability to use GenAI agents in many of them – can make it quite difficult for individual educators to attempt to reduce use and trust that students are doing their own work. 

Teacher Matt Brady, MS, from Atkins Technological and Academic High School in North Carolina, explains a typical scenario. “If a teacher has a quiz on Canvas –  which is our system that we use for content – students can use Copilot [Microsoft’s specialized AI agent]. Copilot is baked into Edge  [Microsoft’s web browser]. That’s a problem. As a result of that, I have a lot of teachers in my building who are going 100% Luddite and collecting phones and using paper and -pencil for everything and testing only on paper-pencil.”

While AI has been quickly adopted for use in classrooms in North America, teachers have become increasingly negative towards AI in schools, according to Dubé’s research. 

“When Generative AI first launched, according to initial studies just seven months after it came out, the majority of K-12 teachers had tried Generative AI, and 80% of them thought it was going to have an overly positive effect on education. They were more positive than students were,” says Dubé. “But now it’s gotten much worse. In fact, only 7% of teachers in a recent Pew study believe that AI is going to have more benefits than harms.”

High school teachers – with students who are now using Generative AI products at high rates –  have a very highly negative outlook on AI now, says Dubé. “80% of them think that AI is going to have more harms or at least as many harms as good outcomes. 64% of middle school teachers have a negative attitude as well towards generative AI.”  

Why are high school teachers so negative about AI in classrooms since its wide adoption? Dubé posits that “it’s likely because they’re being faced with the reality of their students using these systems and they’re seeing the consequences and the difficulties that have arisen from both them and their students using these systems in sort of an unregulated, unfettered way.”

Recent studies have found that when an AI chatbot is readily available in a classroom, some students may rely on it to offload some of their thinking, potentially reducing opportunities for the kind of productive struggle known to support learning, says Ying Xu, PhD, Assistant Professor of AI in Learning and Education at Harvard University.

Recent research on adults using AI tools are finding similar results. One study found that knowledge workers regularly using AI tools reported a sharp decrease in use of their “brain power” and critical thinking skills when using AI, says Naomi Baron, PhD, Professor Emerita of Linguistics at American University and author of Reader Bot: What Happens When AI Reads and Why It Matters. “What the authors of the study said in their conclusion is: without regular practice cognitive abilities can deteriorate over time.” 

A different study from June 2025 compared the brain activity of people of various ages writing an essay either using ChatGPT, standard Google search, or nothing besides their own brain, notes Baron. The study found that the students who were writing solely using their own brains had more electrical activity in their brains at the time of writing, than did the people who used a search tool, than did the people who were using AI, she says. 

Students themselves perceived a difference in the writing they produced under the different conditions in this study. When asked, “How did you feel about the writing?”, those who were doing their own writing without any of these tools felt more ownership over what they had written than people who were using search. Individuals who used ChatGPT felt the least sense of ownership, says Baron.  Similarly, other recent studies show college students worry about using thinking, analysis, and writing skills due to AI use, she notes.

Youth using AI tools as resources for classroom projects are often fed misinformation by the AI tools. “The bots can be very confident – there are some really big issues around misinformation,” says Amanda Bickerstaff, MS Ed, Co-Founder and CEO, AI for Education.

“A lot of the research right now is telling us that generative AI makes mistakes. It hallucinates. It’s got inaccuracies. Now, would you use a textbook if you knew that it actually had significant mistakes in it? Maybe these systems will get better, but the research is showing that these systems and hallucinations are actually increasing as they get more sophisticated. I don’t believe they should be used right now as a resource of information,” says Dubé. 

Using AI as a “tutor” to select, explain, and discuss information with students is risky in that the tutors can be confidently providing guidance based on inaccurate information. Relying on these tools as a primary source for information in this way not only promotes incorrect information, but may also be interfering with students’ abilities to critically evaluate and locate credible sources for factual information.

Awareness that AI might be providing incorrect information is likely not enough to help children critically evaluate its outputs, says Xu. “Children might not be able to effectively translate their awareness into an accurate evaluation of AI-provided information.” Without the skills to know why a piece of information could be wrong, the information can still be mis-used, notes Xu.

It’s common for a suite of AI tools like Google Classroom with Google Gemini to suddenly show up in individual classrooms as part of a school-wide or school district-wide implementation on the nearly-ubiquitous Google Chromebooks students are issued by schools. “All of a sudden all the teachers just have to adopt it because it’s there,” says Dubé. This disempowers educators to make student-centered decisions on how to implement AI thoughtfully based on their expertise as teachers. “I think we should resist that,” says Dubé. “These decisions should come from educators themselves and enable them to make decisions about the tools in their classrooms.”

“A lot of technology that we purchase for classrooms doesn’t really work,” says Dubé.  Many technologies used in classrooms today were not designed for educational purposes. Schools are not requiring strong evidence that the tools work for learning before buying them, and are not evaluating educational technologies effectiveness for actual learning once they are already adopted, he says.

“With generative AI in education, we can’t assume that it’s going to fix education. We have to demand evidence that it will work and that it is working once we buy it,” says Dubé.

Evaluating AI systems for efficacy requires more transparency from technology companies and more critical evaluation by the schools implementing them – neither of which are currently happening today, says Dubé. “We need to evaluate if it’s working and we need to know what it’s actually doing. And those are things that seem very obvious, but they’re actually things that don’t happen,” he says.  “A technology company will come in, and they’ll pitch a product to a school, and they’ll say, ‘It’s powered by AI.’” without disclosing how. On the other hand, schools aren’t asking important questions or requiring evidence and independent research – instead of tech company’s internal testing – that demonstrates that the products are effective for learning. “These are questions that we have to ask edtech companies trying to deploy products into public schools with public funds,” says Dubé.

AI companies have received billions of dollars of funding for creating GenAI tools and are under pressure to recognize profits as soon as possible. Selling their products to school districts is one of their primary strategies right now, says Dubé, despite the fact that educators have not been involved with the development of these systems. 

“Unfortunately what we’re actually getting right now with AI, is a ‘deploy first/test later’ mindset,” says Dubé.  Testing AI tools for educational efficacy should be happening during development and while they are deployed – but this is not currently happening. Instead, students are being used as guinea pigs with powerful technology.  

A central issue is that schools have historically been very poor at evaluating educational technologies in general once they are deployed in schools, because doing so is difficult, notes Dubé.  “It‘s really hard for a school to create a program evaluation system, to systematically evaluate what’s working and what’s not working. Schools are complex, schools are messy, and schools haven’t been empowered or given the resources to actually get it done.” Purchasing educational technology tools, hoping they solve the problem, and then not testing to see if it actually does is unfortunately common, he says.

Teacher training on the AI systems used in their classrooms should be focused on improving how critically they use them, how to selectively use these systems (if at all), and empowering them to make those decisions, says Dubé. 

Unfortunately, the majority of teacher training for implementing AI in schools tends to be purely focused on the nuts and bolts of how to use basic aspects of the tools, such as “this is how you use a prompt,” says Dubé. He notes that most of the training (and money for the training) is coming from the tech companies that make the products, which is probably not the best source of critical training on the use of AI systems. “What we need going forward is for teachers to be taught about these systems, but they need to be taught so that they can be a critical judge of when it’s good to use these things and not good to use these things” in order to better guide their students through the use of these powerful tools, says Dubé.

Most elementary school-aged children are not using GenAI products currently, which is reflected in a greater indifference to AI from elementary school teachers, who feel it isn’t affecting their classrooms overall, says Dubé.

However, the math and reading apps commonly used in today’s elementary school classrooms are  likely to have Generative AI included in the near future, says Dubé, and this indifference may mean more GenAI tools will sneak into younger children’s classrooms. A recent experiment by Dubé’s team presented 200 elementary teachers with several learning app choices for their classrooms, half of which contained GenAI,  and found that over half the elementary educators didn’t care if Generative AI was present in the app or not.  “What that means to me is that going forward at the elementary level, we’re going to have teachers that are choosing apps for their students that have Generative AI running in the background. These things are going to make their way into classrooms – even though the kids aren’t choosing the apps and aren’t using Generative AI for learning, teachers will be choosing apps that are powered by Generative AI because they’re somewhat indifferent,” warns Dubé.

The designers of AI systems that are used in school intentionally collect a lot of information on students, justified as needed to provide maximum AI personalized experiences for students. “That’s actually a real concern when it comes to AI systems,” says Dubé. “They collect not just what the student does on the device, but where the student is, what time, and what other devices are near it. There’s so much information that could be collected.”

The security of this child-specific data is often a “black box” to educators and school districts, and poses a risk to children. “AI doesn’t only have access to a child’s age, where they come from, but it also has access to how a child thinks, what interests them, their preferences, what is convincing to them, and what isn’t. We really need to have systems that respect their privacy and where the decisions that can and cannot be made based on this data is explicit and transparent,” says Mathilde Cerioli, PhD, Chief Scientific Officer at everyone.AI.

While AI can be effectively used as a “mind tool,” where it gets users to reflect and think deeper on their own work, it can be difficult to make sure students are using it this way instead of just passively accepting an AI tool’s redo of their work, warns Dubé.

“You could have a student that writes an essay, writes a paragraph, puts it into generative AI, and then asks it to give them feedback. And then the question here, is when we use these systems and it gives us feedback, what makes it so that we actually listen to the feedback and then do the hard work of reflecting and thinking about ‘What makes me a good writer? Why was the sentence not well written?’ – instead of just accepting changes,” says Dubé.  “There’s not really a lot of thinking when you’re doing a spell check system right now. You just kind of accept all the recommendations. We want to make sure that if there’s a generative AI tool that’s supposed to help students think, it’s not just telling them what to think and they’re mindlessly accepting it. It should be prompting them with questions about how to think and making them think deeper. I have not seen any system that exists commercially that actually does that at this point.”

Research indicates children tend to trust technology that they are familiar with, says Judith Danovitch, PhD, Professor of Psychological and Brain Sciences, University of Louisville, citing research showing four year-old children trusted a Google search engine over a new search engine with the same capabilities. “A lot of four-year-olds haven’t necessarily used Google themselves, just the fact that it’s out there and it’s in their environment makes them seem to be more trusting of it.”

Constant exposure to AI systems in childhood will likely increase trust of these systems by youth due to this familiarity, rather than promoting critical thinking on the information they are providing.

The conversational and human-like nature of GenAI tool responses also makes children more likely to trust them, notes Danovitch. “This conversational nature is very likely to result in children forming what psychologists call parasocial relationships. Parasocial relationships are things that seem like a real social relationship, but they really only go one way…and the stronger that kind of relationship, the more children are going to be likely to want to interact with these things, and to listen to them.”

“Our research found that from a child’s perspective, AI may be seen not just as a tool, but also as a social partner,” says Xu. “On the one hand, it might encourage children’s engagement and model learning strategies. But on the other hand, it could lead children to over-trust AI, as if it came from a trusted human source.”

Many popular apps include “bells and whistles” like animations and sounds that capture the attention of young learners. Dubé and his team looked at different groups of kids – ones who were more and less able to generally focus and pay attention – and their interactions with apps with these distracting features versus ones that were centered on learning content.  “Unfortunately what we found was that the kids who have the least ability to pay attention had the biggest problems when using these educational apps that had distracting content,” he says “When they were using these apps, their attention was being taken away by the animations. It was being taken away by the art in the background that had nothing to do with learning. In fact, they were actually spending more time engaging with all these distracting spots in the application than they were engaging with the focused learning content.” Children who were already able to focus well were much more capable of ignoring these distractible elements.  

Educational AI tools designed with these distractible elements are not working or designed for students already struggling with attention and focus such as children with ADHD. “What we recommended is that we need to make sure that the design of these apps actually have the ability of the child in mind. You need to pay attention to principles of multimedia learning theory… You want to make sure that it’s built for the cognitive system of a young child–and the ones that need it the most, not just the ones that are already doing well,” says Dubé.

The evidence on AI efficacy in the classroom is mixed, says Dubé. While a large number of studies show positive results, an equal number of studies show negative, none, or mixed results, he says. “What that means is that there isn’t a clear answer about whether or not generative AI “works” in an educational setting. Instead, educators have to make informed decisions for themselves, for their classrooms, for their students.”

Most of the research showing that devices are beneficial for learning are compared to poor teaching methods instead of highly effective teachers, notes Abraham Flanigan, PhD, Associate Professor of Educational Psychology, Georgia Southern University.

What about impacts on cognitive skills more broadly? “I don’t think we have really concrete evidence suggesting if using AI would hinder or support children’s cognitive capacities yet,” says Xu. “Most of our research around AI and children’s learning has been focused on short-term learning outcomes. AI tutors can be quite effective to teach children  specific concepts and skills. But we don’t know how this kind of learning of specific concepts and skills would be retained and, in the long run, actually be translated into broader cognitive abilities.”

Most of the research that has been done to date on youth and AI tools has involved adolescents and university students, notes Dubé. “There’s almost no research about how young children are learning with these systems and cognitively offloading.”

As AI tools become available in more and more apps used by younger children, a lack of research for AI use by young users raises concerns that these apps could interfere with cognitive development at a key developmental stage.  “AI conversational partners could offer additional learning experiences, or they might undermine  the growth opportunities that stem from interpersonal interactions,” says Xu. “We don’t have definitive answers yet, but we have seen evidence supporting both possibilities.”

“You cannot talk about artificial intelligence systems without talking about bias,” says Bickerstaff. “These tools primarily have been trained on human data, whether it’s the internet for ChatGPT or every stock photo for Midjourney. What happens is that these tools, because they are probability machines, they actually can not only reflect biases in the training data sets but also amplify them. And that can be very impactful,” she warns.

Cerioli provides a troubling example – “Let’s say you have an AI that is grading a child’s paper. You want to make sure it’s actually grading the paper itself, and not inferring things on the student based on the vocabulary used or the stories shared that can help infer socioeconomic information. The risk could be that the AI rather than grading the work itself is attributing a grade based on those inferred information rather than the work itself. We need that transparency to ensure fairness and to make sure that those systems do not amplify bias and stereotype,” she says.

Trust that regulators are ensuring that AI is being used safely and effectively by children and adolescents is misplaced, say many experts.  “Often people who write the regulation are not experts in AI or in children, two very complex subjects that we need to be able to convey clearly to them,” says Cerioli. Ideally educators and child development experts would consult with industry on what beneficial AI actually looks like, when a feature shifts from helpful to potentially harmful or developmentally risky, and continually iterate to improve products as we build more knowledge and gain more perspective on this technology,” says Cerioli.

The point where research provides a definitive answer on how to use AI tools for learning isn’t going to come, says Dubé, who cites 30 years of research and hundreds of meta-analyses on educational digital games showing that they work… sometimes… it depends. Instead, educators should continue to evaluate and critique AI tools as they encounter them, with one of the most critical questions being – how well are they designed? Were educators and educational experts involved in their design?

Children and adolescents are in a sensitive period of development, and educators should continually question whether the AI tools being used by them in their classrooms  are promoting learning or automating tasks and skills teachers want their students to develop and master, says Cerioli.

Recent research with GenAI tools indicates that when these systems are available, they help children and adults learn and write better “in the moment,” says Dubé. “But once they’re gone, we actually didn’t learn the underlying fundamental skills. We didn’t learn how to generate ideas. We didn’t learn how to summarize and deeply read. We didn’t learn how to explain concepts to ourselves. We let these systems do it for us, and they didn’t give us the opportunity to practice doing that ourselves.” When this “effortful practice” doesn’t happen, this results in students being less skilled in the long run because the system is doing the practice work on your behalf, he warns.

Evaluating AI tools for learning – and whether certain uses are acceptable or not –  involves matching the use with the learning objective. If AI use is not interfering with the learning objective, it’s probably OK, says Dubé.  Dubé provides an illustrative example – a student has been tasked with reading a book and then producing a video about the book by their English teacher and uses a GenAI video editor to make the video.  If the learning objective was to read and understand the book, and summarize the book, then using AI for the video creation doesn’t replace the learning work of the unit. If the assignment was for a video class, where the student was to learn how to use video editing tools, then using AI to make the video would be less acceptable.

Students and educators will see a lot more value in using AI tools when students lean into using the tools to provide more information and focus on something they have already invested significant time and effort on,  instead of replacing learning about a topic, says Bickerstaff.  In this form of AI use, human originality can be centered while the student enjoys the support of tools to help focus on what they want to focus on, she suggests.

Repeating cognitive tasks deepens children’s learning and educators should be careful that AI use is not replacing that repetition, says Cerioli. “We need to make sure that the tools we provide are actually helping children gain mastery, which happens through a fair amount of repetition across time and different contexts. AI can encourage the child to repeat and provide more opportunities for learning but should not automate tasks children are still learning.”

Educators and researchers should be exploring how to support children to develop real-time strategies to interrogate and evaluate information received when they use AI, says Xu. This could include:

• • Asking follow-up questions to the AI tool
  • Asking AI to triangulate and provide responses to a parallel question so that students could compare if AI’s responses to the parallel question are consistent with the responses to the previous question
  • Comparing AI’s responses to what students already know, or what teachers have told them, to emphasize that AI-provided information should be triangulated by another source to increase the trustworthiness

(Xu)

“Children need to understand AI, but they also need to understand what it is to be a human, what are the core skills they need to navigate their world. They need to be able to understand themselves and how they function so they have agency in how they want to use those tools,” says Cerioli. Explain to students why they need to learn to write when ChatGPT could do it for them, such as telling them “You need to be able to speak and express yourself in articulate ways, understand an argument and defend your own opinions and why and how to best relate to others, and have a nuanced perception of the world,” she suggests.

A decade of evidence indicates that children and adolescents are not able to learn how to regulate their own device use while in classrooms naturally, says Flanigan. “l think that calls for some adjustments in curriculum and how we teach that skill in schools.”

Children need to develop skills in thinking and coming up with their own ideas without the aid of AI support at all times. “In order to write down something that you’re brainstorming, you first have to be able to think about it. If you’re not able to have things come out of your own head, but need to rely on what you’re finding electronically, you’re in trouble,” says Baron. Educators can ensure AI-free brainstorming and ideation in their individual classrooms with devices put away.

Educators, students, and families need basic AI literacy – beyond just how to use prompts – in order to use AI effectively and safely. True AI literacy will enable students and teachers to use the tools in ways that help them evaluate risk, protect data privacy, consider what should be reserved as “human,” and when use of the tools constitutes overreliance, says Bickerstaff. Many schools are implementing AI without providing tools or training to increase this literacy.

“I want to enable individuals to use AI in three ways. We want them to be safe, we want them to be ethical, and want them to be effective. There needs to be a balance of developing knowledge and skills, but also new mindsets of using technology,” says Bickerstaff.

Educators and schools need an organizing principle when integrating AI into classrooms, says Dubé. “That principle is that we know what good learning looks like, we know what good teaching looks like, and we have to ask how technology is reflecting that back at us.”  Many people make the mistake of asking how technology is transforming teaching and learning, he says. “That’s not what it’s doing. That’s a sales pitch. The goal is to know what good teaching and learning is, and then to use technologies that enable us to do that. When we do that, we’re going to choose better technologies, we’re going to avoid wasting a bunch of money, and we’re going to put at the center of this whole question the students and the teachers that actually do teaching and learning.”

Many mention preparing students for the workforce of the future as a reason for AI implementation in classrooms. The purpose of education is not just to prepare students for the workforce of today, notes Dubé.   Being able to truly communicate with others, whether that is a personal connection of a coworker, requires skills that children need to learn and practice on their own.

Research shows that educators and parents primarily use user ratings in the app store to gauge the quality of an educational app, says Dubé, regardless of the description and actual features of the app.  This is not a good way to select an app, he says. “We’ve done research on the top apps in the app store and there’s no relationship between the user reviews and the quality of the educational app. That’s because when people are reviewing them, they’re reviewing a bunch of different things about it… if their child had a positive experience…. if they think it looks interesting… if the topic aligns with what their child likes to learn, for example. They’re not necessarily evaluating the educational quality.”

Unfortunately, most companies providing educational apps do not provide educational benchmarks in the App Store, notes Dubé.  Claims of “personalization” and “hands-on” are more marketing-speak than helpful indicators of educational quality, he says. “The developers should have to transparently report on [educational benchmarks], but there’s no requirements, they can write wherever they want.”

What is a good way to choose an educational app for use with children? Dubé suggests considering verifying the following to make sure an AI-enabled app is going to help your child learn instead of provide distraction or entertainment: 

• • Is the app actually teaching the skill you want them to learn? Many apps are not specific about the learning content
  • Does the app scaffold the child’s learning? Does it provide supports like if a child gets stuck, it provides help or hints?  A parent can determine this by playing with the app themselves for several minutes.
  • Does the app provide feedback and tell users when the answer is right or wrong?
  • Did the app involve educators and/or researchers in its development? This can often be ascertained by going to the website for the app.

While many think of safety online or with AI in terms of what type of content children are exposed to, the type of interaction children have with AI matters just as much for their safety, says Cerioli. “A chatbot can give perfectly appropriate advice to teens, but if it does so by framing to have lived a similar experience or positioning itself as being the best resource for this, it still represents a risk by positioning itself as a social partner rather than a tool. AI that act in highly relational ways increase the risk of attachment and emotional reliance. Is AI positioning itself as a bounded tool that gives children information or their best friend and confidant? It has to be child-first AI, which means it’s done in their best interest.”

AI that presents itself in a familiar and “best friend” manner and used for social purposes by youth is often preferred by children yet carries risks to social development and child health. For more discussion on the risks of social AI use to youth, see Children and Screens’ tip sheet.

Educators and parents tend to assume that kids are “digital natives” and can figure out how to use AI technologies for learning effectively on their own, but this is often incorrect, says Dubé. “These technologies aren’t designed for how they learn and kids need to be taught how to use technologies effectively for learning.”

Surveys of youth show that they want to be told by educators what is OK and not OK to use AI for, says Baron. It can be extremely difficult to have definitive answers with the rate of technological change, which makes many educators avoid the conversation. Is spell-check OK? Is AI grammar-check OK? Having honest conversations with students is the first step to work through it, suggests Baron, even if your starting point is “I don’t have the answers.” 

AI tools for classrooms tend to gamify a lot in their personalization features, notes Cerioli. However, some children will become interested in the game itself, and the reward they get, moving them farther from experiencing the learning they are supposed to be doing using the technology, she says.

“What’s happened with a lot of digital technologies is they’ve just allowed for the replication of similar types of teaching  – but now with the added distraction factor,” says Melanie Johnson, EdD, President and CEO, Collaborative for Children.

Research shows that game-based learning can work for learning, says Dubé – but only when the learning itself is the game, he notes. Apps that successfully do this often have educators and researchers consulting with the app designers – something the majority of educational technology developers do not do, he says. A successful learning game will help children practice a skill repetitively in an enjoyable way, that requires flexible thinking, strategies, and a potential for failure – “that’s when game-based learning works,” says Dubé.

Children typically don’t develop complex perspective-taking until 8-10 years of age, says Dubé. Educators should wait to allow children texting or speaking with AI until they have these reasoning skills to think about the mind and how AI systems work. “The student has to have a level of understanding and expertise to be able to engage in that type of back-and-forth dialog with these systems. Younger students don’t have the background expertise to be able to do so,” says Dubé.

Parents and caregivers at home can have values-driven conversations with children using AI for schoolwork to convey there is more at stake than cognitive development or task completion, says Baron. “ We’re trying to build the notion that when you read for yourself, when you write for yourself, you’re developing who you are. You’re developing your own voice. You’re expressing what you want to express as opposed to what some algorithm conjured up.”

There is value to relying on one’s own brain and knowledge – intellectually, socially, and as an individual persona building a sense of who we are, what we think, what we believe, and how to evaluate other people, says Baron. Talking about these values and their importance is something any parent or caregiver can do at home.

Most of today’s email and smartphone text messaging applications feature predictive text capabilities that are driven by AI making guesses about what you might write next. Overuse of these capabilities may damage the quality of communication and the quality of the relationship with your recipient, warns Baron.  Studies have found that if an email recipient believes the sender did not write the message themselves, the recipient tends to think that the sender is arrogant, self-centered, and not interested in you as the recipient.  “You’re not caring enough – to put it bluntly – about the other person to say ‘What do I write that’s personal for this person I care about?” she notes.

Even if children may need to use a lot of AI in the future, it’s important that when they need to use their own brains, they’ll have the independent skill base to be able to do so, says Baron. “I ask, ‘What can you read, what can you write if ChatGPT is down?’ – or Claude, or name your favorite large language model. The whole issue becomes, do you have a foundation that you can call upon? And that’s where building that foundation early on, but then continuing to reinforce it matters.”

Parents and caregivers should be explicitly talking to children about AI as early as possible, says Danovitch. “Children learn a great deal from modeling and from parent-child conversations. I believe that we should start these at home, and potentially also in the classroom, very young, a lot younger than I think we do right now.” 

Adults might think their young child doesn’t even know what Google is, notes Danovitch. “They don’t know what it is, but they are being exposed to it. They hear other people talking about it. Parents can get their child to start understanding how it’s used responsibly by narrating their own use, she suggests. “Adults can say things like, ‘I’m going to use this AI to solve this kind of problem because it’s really good at solving that kind of problem,’ or, ‘I’m not using AI here because it’s probably not going to give me the right answer.”

Parents are currently – and appropriately –  quite overwhelmed with recent AI advancements, but simply sitting down together with your child and using AI tools together with them can help naturally lead to important conversations about how they are using AI, says Pilyoung Kim, PhD, Professor in the Department of Psychology  at the University of Denver, and the Director of the Brain, Artificial Intelligence, and Child (BAIC) Center.  This can include not only finding out how AI may give different answers to the same prompt to an adult and child, but also help parents ascertain if a child is forming an emotional reliance on the chatbot that should be monitored, she notes.

Parents can and should be involved with their local parent-teacher association and schools to ask schools and principals what their policies are for AI adoption, says Dubé. This includes asking for transparency and a full explanation of why their school is choosing to adopt certain tools, he notes.