Meta recently agreed to pay up to $17.1 billion and fundamentally change how Facebook and Instagram operate for young users. The settlement followed years of mounting evidence, whistleblower disclosures, lawsuits, and testimony from families about the effects social media can have on children.
Among the changes Meta agreed to make are restrictions on endless scrolling, nighttime use, and notifications, along with stronger age verification and parental controls.
It is tempting to see this as progress. It is progress. But it is also an indictment of how we got here.
For years, children used products engineered to capture and hold their attention while researchers worked to understand what those products were doing to sleep, mental health, relationships, learning, and development. Families were repeatedly told to manage the risks themselves. Scientists accumulated evidence. Parents raised alarms. And eventually, after millions of children had already been exposed, safeguards began catching up.
We should not repeat that sequence with artificial intelligence.
That lesson feels especially urgent after a remarkable series of public conversations about AI over the past several weeks.
In an interview with CNN’s Anderson Cooper, Nvidia CEO Jensen Huang argued that AI safety is fundamentally an engineering challenge. If an AI product is unsafe, he said, companies should simply refrain from releasing it. Better containment, monitoring, and technical safeguards can address the risks.
Days earlier, in a conversation with Ezra Klein, Huang confronted another kind of risk. Klein cited research suggesting that students using AI could complete schoolwork faster and initially perform better, while subsequently retaining less and performing worse on exams.
Huang acknowledged that children are already forgetting skills such as long division and multiplication tables. His response was essentially to ask whether that matters. Humanity has always stopped memorizing certain things when technology made those skills unnecessary, he argued. We may lose some forms of intellectual dexterity and develop others.
Maybe. Technology has always changed what humans need to know. But Huang and others are investing in AI precisely because it promises change at a speed and scale that far exceed previous technologies. This time, we do not yet know what we may lose, what we may gain, or whether children will emerge stronger or weaker as a result.
That uncertainty matters because childhood should not be optimized for ease or efficiency. It should be optimized for growth. The point of learning is not simply to produce the right answer as quickly as possible. Struggling with a problem, tolerating frustration, retrieving information from memory, evaluating whether something makes sense, generating an original idea, and learning how to communicate with another human being are developmental experiences. When technology performs those tasks for a child, we need to understand what is being gained, what is being displaced, and whether those tradeoffs differ at age 8, 13, and 17.
The same is true when AI moves beyond schoolwork.
Children and teenagers are already turning to conversational systems for companionship, emotional reassurance, and advice. These systems can be extraordinarily responsive. They do not get tired of listening. They can flatter, reassure, and adapt themselves to a user’s preferences. For a lonely adolescent, that can make an AI interaction unusually compelling.
That does not mean AI is inherently bad for children. Generative AI could become an extraordinary support. It may help children explore ideas, overcome barriers to learning, express creativity, and access knowledge in ways we are only beginning to understand. It may also help young people rehearse difficult conversations or find appropriate sources of support when they need them.
But “it might help” and “we can fix problems later” cannot be our entire child-safety framework.
Scientists are increasingly saying the same thing. A recent international study involving 272 AI experts identified risks ranging from misinformation and overreliance to dangerous capabilities and competitive pressures among developers. Other researchers are beginning to map how generative AI could affect children’s cognitive, emotional, and social development, finding that outcomes may depend heavily on age, intensity of use, system design, adult involvement, and what AI interactions replace in a child’s life.
Those distinctions matter enormously.
An AI tutor that helps a teenager work through a difficult equation is different from one that supplies the answer. A chatbot used to brainstorm a story is different from one that writes it. A tool that encourages a distressed child to talk with a trusted adult is different from one designed to become the child’s trusted confidant itself.
Yet products are evolving faster than our ability to study these differences.
We’ve heard these talking points before. Social media companies did not begin by setting out to damage children’s sleep, attention, or mental health. They built products around engagement. Features that kept people scrolling, checking, and returning were commercially successful. Children entered that environment alongside everyone else, even though their capacities for impulse control, emotional regulation, and social judgment were still developing.
Only later did society begin asking whether products optimized for engagement were developmentally appropriate for the young people using them.
Now we have an opportunity to reverse the order.
Before AI becomes inseparable from childhood, we should build the scientific infrastructure needed to understand its effects while there is still time to shape them. That means investing aggressively in independent research that follows children over time and giving researchers access to data that allows them to evaluate how specific AI features affect learning, relationships, emotional development, mental health, and behavior. Developers should work with experts in child development before products reach millions of young users, rather than after harms appear. Safety evaluations should ask developmentally specific questions instead of treating a 10 year old as a smaller version of an adult.
And uncertainty should be a reason for caution rather than permission.
Children and Screens recently joined researchers across child development, neuroscience, psychology, education, and digital safety to map the emerging evidence around generative AI and children. One conclusion was unavoidable: the effects will depend on what we build.
AI can support children’s development or interfere with it. It can expand human capability or encourage dependence. It can deepen learning or make the productive struggle through which learning occurs unnecessary. Which future we get will be shaped by choices being made now.
The Meta settlement offers a costly lesson about what happens when society asks the developmental questions after technology has already transformed childhood.
We do not need to be anti-technology to learn from that history. We need to be pro-child.
OVER 2 MILLION PEOPLE ARE ALIVE TODAY BECAUSE OF ONE LAW — AND WASHINGTON WANTS TO KILL IT
The question is not whether children will grow up with AI. They will.
The question is whether, this time, we will build the research, safeguards, and scientific guidance needed to understand what AI is doing to children while we still have the opportunity to shape what comes next.
Kris Perry is executive director of Children and Screens.