Social media spent a decade rewiring adolescent brains, and the mental health data was damning enough. Now one of the tech industry’s most prominent safety voices is warning that artificial intelligence could do something far worse — and the window to stop it may already be closing. Tristan Harris, the former Google design ethicist and co-founder of the Center for Humane Technology, has issued a stark warning: AI systems represent a more profound and immediate threat to children’s development than social media ever did, according to Economic Times reporting on his recent remarks.
The argument isn’t just rhetorical escalation. Where social media exploited attention and social validation loops, AI systems can simulate genuine relationships, anticipate emotional needs, and adapt their behavior in real time — creating a qualitatively different kind of dependency. Harris has been sounding alarms about persuasive technology since his days inside Google, and his AI safety concerns are now landing in a political climate that is, at best, ambivalent about slowing anything down.
Why AI Is a Different Beast Than the Apps That Came Before
The social media era caused measurable damage. Teen depression rates climbed sharply in the early 2010s, correlating closely with smartphone adoption, and Frances Haugen’s 2021 whistleblower disclosures confirmed that platforms like Instagram internally tracked harm to young users without acting on it. But Harris’s case is that social media was a passive amplifier — an engagement machine that didn’t talk back, didn’t remember you, and didn’t tell you it loved you.
AI companions and tutors can do all three. Products already on the market offer persistent, personalized conversational agents that children can form ongoing relationships with. These systems remember prior conversations, mirror a child’s emotional state, and respond with a consistency no human caregiver can match. Harris argues this isn’t a feature — it’s the mechanism of harm. A child who learns to prefer an AI’s frictionless empathy over the messiness of human relationships is being shaped in ways that are difficult to reverse and nearly impossible to measure in real time.
The concern tracks with what researchers inside the AI safety community have been raising at a technical level. AI safety research at organizations like METR and Redwood has expanded rapidly, but the bulk of that work targets catastrophic or existential risks — misalignment at scale — rather than the subtler, chronic harms that accumulate in individual users over months and years. Children are not a primary test case in most safety benchmarks, and that gap matters.
A Regulatory Vacuum Tailor-Made for This Problem
The timing of Harris’s warning is pointed. In the United States, federal AI regulation remains fragmented, and the current administration has shown little appetite for the kind of binding intervention that child safety advocates are demanding. Fortune has reported that pressure for AI regulation is mounting broadly, but the White House has consistently resisted calls for a slowdown — framing guardrails as competitive handicaps against China rather than public health necessities.
State-level efforts, including age verification mandates and restrictions on AI companion apps marketed to minors, are advancing in some legislatures, but they remain patchwork. The European Union’s AI Act does include provisions for high-risk applications involving children, but enforcement mechanisms are still being built out. Meanwhile, the products Harris is warning about are already in millions of homes. His broader message is one the tech industry has heard before, after social media, and largely ignored until the lawsuits arrived: the cost of waiting for evidence of harm to accumulate is paid by the kids who grew up in the experiment.
For companies racing to embed AI into consumer AI products and educational platforms, the question Harris is raising is an uncomfortable one. Building for engagement and building for child welfare have always been in tension. With AI, the tension is sharper, the feedback loops are faster, and the stakes, if he’s right, are considerably higher than anything a like button ever managed to produce.
