The focus on AI safety disasters ignores everyday harms

This is an opinion piece in MIT Technology Review from Sam Hiner of the Young People's Alliance and Emma Leiken of Hopelab's Responsible Technology Youth Power Fund. Their argument is that policy keeps treating AI safety as preventing the worst outcomes, like a chatbot encouraging suicide, while ignoring everyday harms like teens forming unhealthy attachments to AI companions or students outsourcing all their thinking.

They point to a specific pattern: this year Google backed state chatbot bills that mandated basic protections against sexual content and self-harm, but with broad carve-outs that could exempt popular services. Lawmakers got to claim a safety win, while stronger bills aimed at preventing emotional dependence on AI failed. They cite Nebraska as a place this playbook worked.

We think the core point lands. Preventing catastrophe is the floor, not the job. A product can meet every narrow safety rule and still leave people more isolated, more distracted and less practiced at thinking for themselves. Those harms rarely make a lawsuit or a headline, which is exactly why they're easy to legislate around.

The social media comparison is the strongest part of the argument. We spent a decade debating the worst content on platforms while engagement-maximizing feeds quietly took over hours of people's days. The authors note Meta's platforms have been cited in connection with violence in Myanmar, Sri Lanka and Ethiopia. AI companions and writing tools reach further, into how people think and who they turn to. The authors say a third of teens now choose AI companions over people for serious conversations, and they point to MIT research finding weaker brain connectivity in young people using ChatGPT to write essays.

Where we'd push back a little: the evidence on long-term cognitive effects is still early, and the authors acknowledge there's plenty left to learn. Their proposal that schools prove beyond a reasonable doubt that AI improves learning before using it is a very high bar, and it could keep genuinely helpful tutoring and accessibility tools out of classrooms that need them. Useful AI in education deserves a real test, not a presumption of guilt.

Still, the other proposals are reasonable. Banning companions designed to pretend they're conscious or profess love to users seems like an easy line to draw. Independent scrutiny of how these systems affect relationships and thinking is the kind of oversight any mature industry should expect.

Our read: the industry benefits when safety means avoiding disasters, because that framing never asks it to give up engagement. If AI becomes a default companion and default ghostwriter for a generation, the cost won't arrive as one terrible event. It will arrive slowly, as skills and relationships people never built. That's worth regulating before it's normal.

Hiner's organization coordinated advocates who helped pass AI companion laws in eight states in 2026, including California's SB1119. Leiken previously led global youth safety and well-being policy at TikTok. Both are advocates with a stated position, and the piece should be read as argument rather than reporting.