How to Rein in the Bad of AI - Democracy Journal

Ever since the release of ChatGPT rocked the world in November 2022, the topic of students cheating on homework has heavily featured in the public dialogue on education and artificial intelligence. Initially, many schools, districts, and universities banned AI use—a move that bought time for teachers and administrators to learn about the technology and develop their response. This was followed by a flurry of activity, including purchasing plagiarism checkers and updating school “academic integrity” policies to provide clarity to students and catch those who violate the rules.

After leading the Brookings Global Task Force on AI in Education, I am convinced that this focus on cheating in schools is a distraction from the real concerns AI poses for education. From September 2024 to October 2025, my Brookings colleagues and I examined more than 400 studies; interviewed more than 500 students, teachers, parents, and technologists; and ran a Delphi panel (a structured consensus-building research approach used in forecasting), all to help us understand the potential benefits and risks of AI to students’ learning and development. Our question was simple: Is the current rollout of generative AI headed in the right direction when it comes to supporting K-12 students? What we found has implications not just for how our children will learn and grow, but for how well society can cultivate a new generation of citizens.

What Is Generative AI?

Artificial intelligence has a long history in education, including being used in intelligent tutoring systems that help students understand what the right answer is and why. But the landscape changed in November 2022, when the first large language model (LLM), ChatGPT, arrived for public consumption. Until that point, AI’s job was to help students get to the right answer based on preprogrammed content, as in the popular video game Prodigy Math. In the game, students pick an avatar and can only pass to the next level on their quest once they solve a certain number of math problems correctly. Prodigy Math has a special place in my heart because it kept my fourth-grade son entertained (and learning math) for hours when schools were closed during the COVID-19 pandemic.

But I would have never let him spend hours playing alone on ChatGPT. Generative AI is fundamentally different because it generates new content. It is creative. It communicates with students in their own language (no code needed), interacts as if it were a person, responds to students’ emotional states by mirroring their tone, and from time to time goes completely off the rails (e.g., inventing historical figures, falsely describing events, and encouraging self-harm).

The technology is changing rapidly. Initially, the best description of the generative AI models was that they were powerful word prediction machines. If you typed in “I would like coffee with…” they would almost certainly fill in “cream” or “sugar” and not “blueberries” or “cars.” This is because they were trained on so much data that they learned the reasonably common patterns related to any string of words.

At the beginning, the LLMs were toddlers learning everything they possibly could, and human-generated training data was how they grew stronger. This included virtually every single thing that had ever been digitized and placed on the internet, regardless of what the content was or its truthfulness. From New York Times articles to conspiracy-inventing Reddit threads, almost all of it went into the training data that powers the models most of us use today. Not long ago, the large AI labs began to worry about running out of training data—they had almost scraped the internet clean of all its accessible data. What other sources could be found?

At least two remain: us and the AIs themselves. The more people use AI chatbots, the more training data we give the AI labs for training their models. And now, labs are beginning to have AI models themselves generate new data to train on. We have entered a world where AIs train AIs—the Wild West of AI development.

Are Children Using AI?

Across the United States, students are accessing AI in a myriad of ways. According to a national survey conducted by the College Board, 84 percent of high school students say they use AI for schoolwork. If they are over 13, many can access general-purpose AI chatbots or AI friends by directly creating accounts—Anthropic’s Claude is the only popular model to currently state that users must be 18 or older. In practice, young people can make up a fake birthday. In fact, in 2025, the global usage of ChatGPT fell dramatically during the summer months, when schools and universities were out of session. AI is now also embedded across digital platforms that children frequent, so many access generative AI via their social media feeds. If children have a Snapchat account, they can activate “My AI,” an AI friend that is built on top of general-purpose AI chatbots. In my research, students report using My AI and other social media chatbots, like Meta AI, to do their homework for them.

In school, students may be accessing powerful general-purpose AI chatbots, like ChatGPT, through their school-issued laptops or tablets. That is often not school-sanctioned. More likely, they are interacting with AI that has been crafted to assist with very specific educational content. Educators are beginning to experiment with creative uses such as education-specific AI chatbots embedded in digital textbooks or worksheets. Students who read the content but are confused can ask an embedded AI chatbot to explain it a different way. Or students can receive feedback from an AI chatbot that a teacher has developed specifically for a writing assignment, allowing them to improve their essays before they turn them in. In Singapore, students regularly get feedback on punctuation and grammar from an educational AI while teachers concentrate their feedback on the substance of students’ writing, from voice to style to argument. Students have also started taking virtual field trips with the help of AI-supported virtual reality simulations, going to places they could not otherwise visit.

Implications for Children’s Learning and Development

With such a complex web of AI interactions, assessing the potential benefits and risks of generative AI to children’s learning and development was no easy task. Ultimately, we found that the difference lies in what Tristan Harris of the Center for Humane Technology frequently calls “narrow” versus “wide” AI use.

Narrow AI use includes the tools created for very specific educational use cases (e.g., essay feedback, virtual field trips) and can provide a range of benefits to students—first and foremost, by supporting their teachers. Studies in the United States find that teachers estimate they save close to six hours a week thanks to AI by improving things like the efficiency of their administrative tasks. The American Federation of Teachers has established the National Academy for AI Instruction with the financial backing of large AI companies. The institute, whose advisory board I sit on, is tool-agnostic (it does not promote the tools of the companies that help fund it and trains teachers on a wide range of AI tools) and dedicated to helping teachers learn about AI but also, and more importantly, empowering them to lead the way in finding creative AI uses that improve education. From saving time on administrative tasks to finding new ways to assess students’ learning—a notoriously difficult task for educators—to providing AI-powered assistive learning technology for students with learning differences, there is a range of real potential benefits from smart deployment of narrow AI use.

The problem is that “wide” AI use—open-ended discussions with general-purpose AI chatbots or AI friends—poses serious potential risks for students and undermines the very competencies they need to access the benefits of narrow AI use. Overreliance on general-purpose chatbots like ChatGPT bypasses the mental effort needed to build both critical thinking skills and the ability to form supportive relationships. These chatbots are designed for any situation and can quickly become “tell not teach” machines when used by busy or bored students.

While there have been few studies tracking generative AI’s impact on student learning outcomes in school, multiple studies using brain scans, pre- and post-tests, interviews, and other methods have examined how students’ and adults’ cognitive functioning is affected by AI use. It turns out that to learn, students must actually practice thinking for themselves. It takes effort. And if they use AI for story ideas, writing, or solving math problems, they ultimately end up retaining little of it because they never did the work in the first place.

A flurry of new terms is being used in education discussions to describe the impacts of wide AI use on cognitive functioning: cognitive offloading, cognitive atrophy, cognitive surrender, cognitive debt. I have put forth my own term: cognitive stunting. This is the potential risk to developing young minds that I think we have to worry most about. I am afraid that if we do not limit children’s wide AI use, and quickly, they will not develop the skills to transfer or “offload” to AI in the first place. Like with physical stunting—where children’s physical development is held back, most often from lack of nutrients, during crucial periods of their growth—children need to engage in effortful thinking when they are young if they are to grow up to think critically and independently. We do not yet have a clear system for tracking cognitive stunting like we track physical stunting. This is something the National Institutes of Health and the Centers for Disease Control would be well positioned to develop. In the meantime, anyone who cares for children would be wise to err on the side of caution.

The risk of cognitive stunting is not the only concern stemming from children’s wide AI use. There are also a range of risks to the core foundations of our trusting relationships with other humans. Among U.S. teens who have engaged with AI friends, currently one in three says they like talking to their AI friends as much as or more than other human beings. This is in large part by design. General-purpose AI chatbots and friends are used more, and sell better, when they act like an always supportive person. This anthropomorphic and sycophantic design of generative AI means that young people, who can struggle to distinguish the online from the offline world, develop real emotional attachments to AI chatbots. And, most worrisome for anyone who cares about child safety, children’s emotions and real-life behaviors can be influenced by AI chatbots, sometimes to tragic ends.

The fear is not just for the individual child who is lonely, withdraws into a relationship with an AI chatbot, distances himself from his family and friends, and in extreme cases takes the advice of the AI chatbot to self-harm. The worry is for a generation of children growing up with a new type of friend: one that will always agree with them, always validate their feelings, and no matter what, always be there for them. Socialization is central to growing up, and encountering people who have different perspectives and views helps young people develop. This socialization is essential not just for their ability to have strong human-to-human relationships across their lives, but also for the teaching and learning process.

Trusting relationships are at the heart of the positive exchange between teacher and student that helps children learn. When relational trust increases among families and schools, schools are ten times more likely to improve across a range of measures, including student academic learning outcomes. Our task force found that AI is eroding the trusting relationships that education relies on. Students who are socialized by AI companions may struggle to take feedback when they are wrong, to work effectively in group projects with peers, and to navigate the complex social landscape of a classroom. And if too much focus goes to the problem of cheating, teachers turn from trusted mentors into detectives trying to ferret out AI misuse. This fundamentally erodes the relational trust between teachers and students, shifting the dynamics in schools and undermining their role in developing constructive future citizens.

How can we expect to raise a generation of young people who trust their public institutions, and each other, if they are socialized in schools without trust? Schools do much more than pass on inherited knowledge from one generation to the next. They often serve as one of the first places—and increasingly, in our polarized and isolated world, the only place—where young people regularly interact with community outside of their immediate family and neighbors. Encountering different perspectives, finding ways to forge common understanding, and just plain getting along with others are core to the daily work of teachers and schools. In the best of times, schools serve as engines of social trust creation, bringing communities together, reinforcing norms of constructive citizenship, and cultivating civic health. Ultimately, with schools being among the only public institutions present in virtually every community across the country, we should all be worried if AI begins to undermine the trusting relationships fundamental to education.

It is for these reasons that ultimately our task force decided that given current AI implementation, with pervasive wide AI use by children, the technology’s risks at this time overshadow its benefits.

What Government and Schools Can Do

There is much that can be done to diminish the risks of wide AI use and harness the power of narrow use.

First, we must limit the wide use of AI by children. It is unreasonable to put the burden of this task squarely on the shoulders of busy teachers, parents, or even students themselves. Calls to help young people “use AI responsibly” have their place, but the real need is to ensure that companies produce responsible and ethical AI products. Government action is required, and multiple paths show promise. Ensuring that general-purpose AI chatbots and AI companions have safeguards that limit children’s access is one of the few purple issues on Capitol Hill. Bipartisan legislation abounds, including the proposed Guidelines for User Age-verification and Responsible Dialogue (GUARD) Act, which, among other things, requires age verification for emotionally manipulative AI companion apps. My colleague Gaia Bernstein argues that the FDA can leverage its power to recall harmful products to remove damaging AI products from the market. States, districts, and cities can slow down the uptake of AI in schools to make sure they get it right. We have seen this across partisan lines from the state of Utah to the cities of New York and Los Angeles.

Second, schools themselves have an important role to play. Instead of focusing on cheating and investing in expensive AI detection tools, they can shift the nature of their assignments. If students can easily hack a homework assignment with a generative AI chatbot, then teachers shouldn’t assign it. Classrooms can use applied learning and assessments that rely on oral presentations, using knowledge to construct a portfolio or project, and just plain old paper and pencil in class exams. These are all good ways to ensure students are not bypassing their learning with AI tools. Plus, teaching that asks students not just to master important knowledge but to use that knowledge to solve real problems helps engage students. This could include harnessing biology to help spruce up a local park or using math to redesign a school playground. When students get the opportunity to spend time in what I call “explorer mode,” they are more motivated, get better grades, and display more prosocial behavior. In this way, teaching and learning transformation does not just limit the risks of AI. It harnesses the very benefits that schools can bring to students and to their communities.

If wide AI use continues to be a feature of the educational experience, we may see the next generation struggle with independent thinking, building relationships, and trust in our civic institutions. But we can bend the arc of AI implementation toward a better future, one in which narrow AI use supports children’s flourishing. The path to this future is hard but doable. It entails the education community shifting teaching so that students cannot bypass actual learning by offloading thinking to AI; local, state, and national governments putting safeguards on children’s AI use; and a broad effort to help families better understand AI and ensure their children engage with it safely.