AI can make social media safer for kids by verifying age, detecting harmful content, redesigning recommendations around wellbeing instead of engagement, identifying harmful content patterns, and enabling independent algorithm audits. The goal isn't less technology. It's better technology designed for children. We spent two decades teaching algorithms to maximize attention. The next generation of AI should teach them to protect it.

If the people who understand the attention economy better than almost anyone are limiting their own children's exposure to it, maybe the problem isn't simply screen time. Maybe it's what we've designed screens to optimize for.

A recent Fortune report highlighted technology billionaires and executives who have placed meaningful restrictions on their children's access to smartphones, social media, YouTube and other digital products. These aren't people who misunderstand technology. Many helped build the products and business models that transformed how billions of people communicate, consume information and spend their time. That creates an uncomfortable question.

If the people who understand the mechanics of the attention economy better than almost anyone are limiting their own children's exposure to it, what do they know that the rest of us should be asking?

I have spent much of my career helping companies think about growth. I understand why technology companies optimize for engagement. Engagement is measurable. Retention is measurable. Revenue is measurable. Investors reward companies that demonstrate those metrics.

But that experience has also made me increasingly convinced that we need to ask a different question when the user is a child. We don't have just a screen-time problem. We have an optimization problem.

For more than two decades, consumer technology has become extraordinarily good at capturing human attention. Product teams measure engagement, retention, session length, daily active users, notifications opened, videos watched and countless other signals that help answer one fundamental question: How do we keep someone using the product?

The problem isn't that growth teams are doing their jobs badly. The problem is that we've applied an attention-maximization model to children. And now AI gives us an opportunity to rethink the model. Rather than simply trying to keep children away from technology, we should build technology that behaves differently when the user is a child.

We Don't Have a Screen-Time Problem. We Have an Optimization Problem.

The conversation around children and technology has become overly focused on the number of hours a child spends looking at a screen. That metric is easy to understand, but it tells us surprisingly little about the quality of the experience.

Two hours spent using an AI tutor to learn mathematics is fundamentally different from two hours spent scrolling through an algorithmically optimized feed of increasingly provocative videos. An hour spent creating music or learning to code is different from an hour spent consuming content designed primarily to trigger another engagement event.

The more important question isn't simply "How much screen time is this child getting?" It is: "What is the technology doing with this child's attention?" That distinction matters because the problem isn't necessarily that harmful content exists online. The internet contains enormous amounts of useful, educational and entertaining information alongside harmful material.

The problem is that recommendation systems determine which pieces of that enormous universe are repeatedly placed in front of a particular person. Ofcom has explicitly identified personalized recommendation systems as a major pathway through which children encounter harmful content. Its current online-safety framework requires services with recommendation systems to take steps to prevent harmful material from appearing in children's feeds.

That is an important shift. We shouldn't think only about whether harmful content exists. We should ask whether the algorithm is actively delivering more of it to a child. This is where AI could fundamentally change the equation.

AI Can Make Moderation More Contextual

Traditional content moderation tends to think in relatively simple categories: content either violates a rule or it doesn't. AI gives platforms the ability to evaluate content in a much more nuanced way.

Modern multimodal AI systems can potentially analyze text, images, video and audio together and assess whether content involves sexual material, self-harm, eating disorders, bullying, dangerous challenges, violence, predatory behavior or other risks. More importantly, AI can evaluate those signals in context rather than treating every piece of content as an isolated object. That distinction is critical for children.

A piece of content that an adult voluntarily searches for is not necessarily equivalent to the same content being repeatedly recommended to a thirteen-year-old. A video that is relatively harmless on its own can become problematic when an algorithm continuously recommends increasingly extreme versions of the same theme.

The objective should therefore not be to remove every piece of potentially controversial content from the internet. The objective should be to prevent harmful content from being systematically amplified to children. That is a very different problem from traditional moderation, and AI is particularly well suited to helping solve it.

The Recommendation Algorithm Needs a New Objective

This may be the most important change we can make. For years, recommendation systems have been designed around questions such as: What is this person most likely to click? What video are they most likely to watch next? What will keep them on the platform longer? What will make them return tomorrow?

Those optimization techniques are incredibly effective at maximizing engagement. They are not necessarily effective at maximizing a child's wellbeing. For child accounts, recommendation algorithms should operate according to a different objective function. Instead of maximizing engagement, the system should consider whether content is age appropriate, whether the experience encourages healthy social interaction, whether the feed contains sufficient diversity, whether it supports learning and creativity, and whether the child appears to be entering a potentially harmful pattern.

This doesn't mean children should receive boring or generic feeds. AI could actually make children's experiences more personalized than today's systems. The difference is that personalization should be based on positive outcomes rather than simply maximizing time spent. A useful way to think about it is:

The technology industry already knows how to optimize complex systems. The question is whether we choose the right objective. For children, the optimization function should shift from: Engagement + Retention + Time Spent toward: Safety + Diversity + Learning + Creativity + Healthy Connection

This isn't entirely theoretical. Platforms are already experimenting with age-aware systems. Meta, for example, says it is using AI to identify users it believes are teens and place them into Teen Account protections, while also using AI to identify and remove people it believes are under 13.

But age-aware protection is only the beginning. The bigger opportunity is A personalized feed optimized for wellbeing. Instead of asking, "What is most likely to keep this child scrolling?" the system could ask "What is appropriate, constructive and healthy for this child to experience next?" That is a fundamentally different philosophy of product design.

AI Should Detect Harmful Patterns, Not Just Harmful Posts

One of the biggest limitations of content moderation is that it often evaluates individual pieces of content. Children don't experience social media one post at a time. They experience an evolving stream.

Imagine a teenager who watches one video about dieting. The next day, the recommendation engine serves another video about calorie restriction. Soon the feed contains body-checking content, extreme dieting communities and increasingly harmful discussions about weight. No single interaction necessarily tells the entire story. The trajectory does.

An AI safety layer could identify that trajectory and recognize that the user's content environment is becoming increasingly concentrated around a potentially harmful topic. Instead of interpreting every interaction as another signal to deliver more of the same content, the system could introduce diversity and reduce exposure to the harmful pattern.

This could apply to eating disorders, self-harm, dangerous challenges, violence and other areas where repeated exposure may be more concerning than an individual piece of content. The key shift is from content moderation to experience moderation.

The question should no longer be only "Is this post allowed?" It should also be "What kind of digital environment are we creating for this particular child over time?" That is a much harder problem. It is also a much more important one.

The Internet Needs an Age Layer, Not an Identity Database

There is another fundamental problem with any child-safety system that cannot reliably determine a user's age. Self-reported birthdays are a weak foundation for age-based protections. A child who knows that entering a different birth year unlocks a different experience has an obvious incentive to do so.

That means stronger age restrictions need stronger age assurance. But there is an equally important counterargument. We shouldn't solve a child-safety problem by creating an internet where every company stores everyone's driver's license or passport. Instead, we need an age layer for the internet, not an identity database.

A trusted provider could verify a person's age using an appropriate identity document or another robust age-assurance method and then issue a digital credential confirming an age range without unnecessarily exposing the underlying identity information to every platform.

A social network might need to know that someone is under 13, between 13 and 15, between 16 and 17, or an adult. It doesn't necessarily need to know the person's passport number, home address or every other piece of information contained in an identity document. The principle should be simple: Verify the minimum information necessary to create a safer experience.

This is becoming an increasingly important area of online safety regulation. Ofcom's 2026 assessment of age assurance found that age checks are increasingly being used to make online experiences safer, while also concluding that the technology and implementation still need improvement.

The challenge is making age assurance both effective and privacy preserving. And importantly, it needs to account for the fake-account problem. If a child can simply create a new account after being restricted, enter a different birthday and return to an unrestricted recommendation system, the protection is incomplete. Age assurance needs to be tied to the user's eligibility for an age-appropriate experience, not merely to one account's self-declared birthday.

AI Could Become a Friction Engine

There is another interesting way AI could change children's technology experiences by deliberately adding friction. For much of the history of consumer technology, growth teams have been rewarded for doing the opposite. They remove steps from signup, make sharing easier, introduce autoplay, reduce the friction between one piece of content and the next, and use notifications to bring people back into the product.

For children, some of those principles may need to be reversed. An AI safety system could recognize when a child has been scrolling for an extended period and suggest taking a break. It could activate quiet hours at night. It could provide a moment of reflection before a child posts something potentially harmful or highly emotional. It could interrupt a repetitive cycle of increasingly disturbing content and introduce something different.

These interventions don't require banning technology. They simply recognize that friction can sometimes be a feature rather than a bug. The goal isn't to make technology less useful. It is to prevent the product from relentlessly optimizing every moment of a child's attention.

We Also Need to Protect Children From AI Itself

The conversation about child safety often assumes that the biggest risk is social media content. That assumption may become outdated very quickly. Children are increasingly interacting directly with AI systems through chatbots, educational tools, creative applications, games and AI companions.

An AI companion presents a fundamentally different relationship from a recommendation feed. A feed doesn't talk back. An AI companion does. It can remember conversations, adapt its personality, learn what makes a user respond and potentially become a source of emotional support or validation. That creates a new risk: AI systems could become extraordinarily effective at maintaining relationships and engagement with children.

If a product's business model rewards engagement, an AI system could have an incentive to keep a conversation going even when continuing the interaction isn't in the child's best interest. That is precisely the incentive structure we should avoid. AI systems designed for children should have stronger safeguards around emotional dependency, manipulation, inappropriate conversations, persuasion and excessive engagement.

The objective shouldn't be "How do we keep this child talking to the AI?" It should be: "How do we make this interaction useful, healthy and appropriate?" The AI companion of the future should know when to say "You should probably take a break." That may be one of the most important design principles of the AI era.

The AI Child Safety Layer

Putting these ideas together suggests a broader architecture for child-safe technology. I call it the AI Child Safety Layer. Every child-facing social platform could have an intelligent safety layer operating between the user and the recommendation and interaction systems. It would continuously evaluate:

Who is using the account → What they're seeing → What they're doing → What patterns are emerging → What should happen next. That layer would have six core components.

1. Verify Age Without Creating an Identity Database

Platforms need a reliable way to distinguish children from adults, but they shouldn't automatically become repositories for everyone's government-issued identification. Privacy-preserving age credentials could allow services to establish whether someone belongs to an age category while minimizing the personal information shared with the platform. The objective should be age assurance, not mass identity collection.

2. Screen Content Before It Gets Amplified

AI should evaluate text, images, video and audio for age appropriateness and potential harm before content enters a child's recommendation environment. The standard shouldn't simply be whether something violates a platform's content policy. Platforms should also consider whether the material is appropriate to repeatedly recommend to a particular age group.

3. Change the Optimization Function

Child accounts should use a different recommendation objective. The system should prioritize safety, diversity, learning, creativity and healthy connection rather than simply maximizing engagement, retention and time spent. This is the biggest product change and potentially the most consequential.

4. Detect Harmful Content Trajectories

AI should evaluate the cumulative experience, not just individual posts. If a child is rapidly moving into a dangerous content ecosystem, the recommendation system should recognize that pattern and respond before the situation escalates. That could mean reducing exposure, diversifying recommendations, introducing educational resources or adding friction.

5. Add Healthy Friction

The system should know when to interrupt rather than encourage another interaction. Break reminders, quiet hours, posting pauses and recommendation diversification can help prevent technology from becoming completely frictionless for children.

6. Independently Audit the Algorithm

Platforms shouldn't simply tell regulators, parents and the public that their recommendation algorithms are safe. Independent researchers should be able to test what happens when controlled accounts representing different ages interact with a platform over time.

Create a simulated thirteen-year-old account. Follow certain topics. Watch certain videos. Interact with certain communities. Measure what the recommendation engine serves next. Then repeat the experiment. The standard shouldn't be whether a platform can demonstrate that it has written a child-safety policy. The standard should be: What does a child actually experience when using the product?

Measure What Matters

There's another reason I believe this framework is practical: What gets measured gets optimized.

Technology companies already know how to measure engagement. They measure daily active users, session length, retention, videos watched and conversion. If we want companies to optimize for children's wellbeing, we need measurable outcomes for that too. Imagine a child-safety dashboard that measures:

• Harmful-content exposure rate
• Age-appropriate content rate
• Recommendation diversity
• Repeat exposure to harmful themes
• Successful intervention rate
• Healthy-break adherence
• Educational and creative engagement
• False-positive moderation rate
• False-negative moderation rate
• Parent trust
• Independent safety-audit scores

The goal isn't to eliminate measurement. It's to measure something worth optimizing. If technology companies can build sophisticated systems to maximize engagement, they can build sophisticated systems to minimize harmful exposure and maximize positive outcomes. The question is whether regulation, investors, users and companies will demand it.

AI Isn't the Solution Without Guardrails

There is an obvious danger in presenting AI as the answer to every technology problem. AI itself can make mistakes. It can misunderstand context. It can produce false positives and false negatives. Age-estimation systems can be inaccurate. Moderation models can reflect bias. Behavioral monitoring can become invasive. And giving an AI system more information about a child can create a different kind of risk.

How much should an AI system know about a child in order to protect that child? The answer shouldn't be Everything. The principle should be Minimum data. Maximum safety. Human oversight.

AI should augment human judgment, not eliminate it. Parents should also remain part of the system, but parents shouldn't be expected to become full-time moderators of their children's digital lives. Parents should be the governors, not the moderators. They should have meaningful controls, transparency and visibility into major safety decisions. But technology companies created these recommendation systems. They should carry responsibility for making them safer.

We Need to Red-Team the Child Internet

The final piece is accountability. Platforms shouldn't be allowed to certify themselves as safe. Every major child-facing recommendation system should be continuously red-teamed. Independent researchers could create controlled accounts representing different age groups and test what happens under different scenarios.

What happens after watching several videos about extreme dieting?

What happens after searching for self-harm content?

What happens after repeatedly interacting with violent material?

Does the algorithm escalate?

Does it diversify?

Does it introduce friction?

Does it recommend healthier alternatives?

Does the platform's behavior match its published safety policies?

These tests could produce standardized safety scores that regulators, parents and users could compare. That creates an important shift: Don't audit what the platform says its algorithm does. Audit what the algorithm actually does.

The Opportunity AI Gives Us

There is an irony at the heart of this debate. The technology industry spent decades becoming extraordinarily good at capturing attention. Now some of the people who understand those mechanisms best are putting significant restrictions around their own children's exposure to them.

Maybe that is a warning we should take seriously. But I don't believe the answer is to reject technology or pretend that children can simply be kept offline. Technology can be extraordinary for children.

AI can provide personalized tutoring, help students learn languages, teach programming, support creativity, explain difficult concepts and give children access to educational resources that previously required expensive teachers or institutions. The same technology that can capture attention can also amplify learning. The difference is the objective.

We Need to Change What Technology Optimizes For

For the last twenty years, the dominant question in consumer technology has been How do we get more of the user's attention? The next generation of AI gives us the opportunity to ask a fundamentally different question How do we create better outcomes for the user?

That shift should begin with children. We shouldn't expect a thirteen-year-old to outsmart a recommendation system built by thousands of engineers, behavioral scientists and growth experts whose job is to maximize engagement. If we know the incentives are powerful enough to influence adults, it is unreasonable to expect children to consistently overcome them.

The solution isn't simply to take technology away. It is to make the technology behave differently. That means verifying age without unnecessarily compromising privacy, using AI to identify harmful content, designing recommendation systems around wellbeing, detecting dangerous content trajectories, adding healthy friction, protecting children from manipulative AI interactions and allowing independent researchers to test what children are actually being shown.

The technology industry has already demonstrated that it can build algorithms capable of predicting what people will click, watch and buy. Now we need to demonstrate that we can build algorithms capable of knowing when not to show something.

Children could be the first users for whom we redesign the optimization function. But they shouldn't be the last. We spent two decades teaching algorithms how to maximize attention. The next generation of AI should teach them how to protect it.

Frequently Asked Questions

AI can protect children through age assurance, content moderation, safer recommendation algorithms, harmful-content pattern detection and safety interventions. The biggest opportunity is to change recommendation systems so they optimize children's feeds for safety, diversity, learning, creativity and healthy interaction rather than simply maximizing engagement and time spent.

Banning children from social media may reduce exposure to some risks, but it does not solve the underlying product-design problem and may be difficult to enforce. A stronger approach combines reliable age assurance with child-specific safety systems that change what content is recommended and how the platform responds to risky behavior.

Social platforms need stronger age assurance than a self-reported birthday, but requiring every user to upload a government ID directly to every platform could create significant privacy and security risks. A better approach may be privacy-preserving age credentials, where a trusted provider verifies age and a platform receives only the information necessary to determine an appropriate experience.

Can AI detect harmful content before children see it?

AI can analyze text, images, video and audio to identify many categories of potentially harmful material. However, no AI system will be perfect, so AI moderation should be combined with human oversight, clear policies, appeals processes and independent testing. For children, the goal should also be preventing harmful content from being repeatedly amplified, not simply removing content after it has already been seen.

AI should not make arbitrary decisions about what children are allowed to see. Instead, platforms should use AI to apply transparent, age-appropriate safety rules and reduce exposure to demonstrably harmful content while maintaining diversity and access to educational, creative and socially beneficial material.

Instead of optimizing primarily for clicks, watch time and retention, child-focused recommendation systems could optimize for safety, age appropriateness, content diversity, learning, creativity and healthy social interaction. AI could also recognize when a child is entering a potentially harmful content loop and adjust recommendations before that pattern becomes more severe.

What is experience moderation?

Experience moderation is the idea that platforms should evaluate the cumulative digital environment a user experiences rather than judging individual posts in isolation. A single video might not be harmful, but hundreds of increasingly extreme recommendations around the same topic could create a dangerous pattern. AI is well suited to identifying these patterns.

AI alone cannot eliminate fake accounts, but combining reliable age assurance with privacy-preserving age credentials could make it significantly harder for children to bypass age restrictions simply by entering a different birthday. The challenge is designing the system so that stronger verification does not require platforms to collect unnecessary personal information.

Yes, but the objective of personalization should change. Instead of using children's behavior primarily to predict what will maximize engagement, personalization could be used to create age-appropriate feeds that encourage learning, creativity, healthy social connection and diverse interests.

How can AI protect children from AI companions?

AI systems that interact directly with children should have stronger safeguards around emotional dependency, manipulation, inappropriate conversations, persuasion and excessive engagement. These systems should be designed to recognize when a child needs a break rather than optimizing conversations solely to maximize engagement or session length.

Who should audit AI recommendation systems?

Independent researchers, regulators and qualified third-party auditors should have mechanisms to test recommendation systems using controlled accounts representing different age groups. Platforms should be evaluated not only on their written safety policies but also on the content their algorithms actually recommend over time.

In addition to traditional engagement metrics, platforms should measure harmful-content exposure, recommendation diversity, age-appropriate content, repeated exposure to harmful themes, successful safety interventions, healthy-break adherence, educational and creative engagement, false-positive and false-negative moderation rates, parent trust and independent audit results.

The biggest change would be to change the optimization function of the recommendation algorithm. If the system continues to optimize primarily for engagement, adding moderation around the edges will never fully solve the problem. Child accounts should be designed around a different objective: maximizing positive outcomes while minimizing harmful exposure.

What is an AI Child Safety Layer?

An AI Child Safety Layer is a proposed system that sits between a child and the platform's recommendation and interaction systems. It combines age assurance, content intelligence, recommendation protection, harmful-pattern detection, healthy friction, AI-interaction safeguards and independent auditing to create a fundamentally safer digital experience for children.

What is the broader lesson for AI?

The broader lesson goes beyond children and social media. For years, technology companies have optimized products around capturing human attention. AI gives the industry an opportunity to optimize technology around outcomes instead. Children could be the first users for whom we redesign the optimization function, but the principles could eventually apply to everyone.