We’ve all heard the warnings for years. Screens are harming our kids’ brains, and it seems like it’s coming to a point of urgency for teachers, administrators, and parents.
Test scores have stalled and attention spans appear shorter—although I’m always a little suspicious of how we measure that. As part of Congressional testimony, a researcher posits that Gen Z may be the first generation in modern history to underperform their parents on cognitive evaluations. In his testimony, Dr. Jared Cooney Horvath spoke about classrooms full of digital distractions. Today’s children spend more time with technology than any generation, and somehow literacy and problem-solving keep disappointing us.
Yes, the concern is real. But the conversation keeps focusing on a distinction that may matter more than screen time itself—what the screen is actually asking the child to do.
What Sesame Street Already Taught Us
We’ve run this experiment before, just with a different box.
During the television era, researchers found that the effects of screen exposure depended less on the hours logged than on what a child was doing during them—context over content. “Screen time” like Sesame Street produced measurable gains, while passive viewing and background television correlated with weaker language development and shorter attention.
Same medium. Different demand on the viewer. Different outcome. I want to say it’s that simple, and for television it might have been.
Where the Analogy Breaks
Sesame Street could prompt a child. It could ask a question, leave a pause, invite a guess. What it couldn’t, a key point of differentiation, was to answer back. The participation it demanded was real because the show had no idea what the child said in response. The screen never adjusted.
But today, a large language model doesn’t just adjust, it iterates. And along the way, this unique form of engagement changes the cognitive geometry of the box to a more “learner centric” dynamic. And that changes what “interactive” even means, in a way I don’t think we’ve fully reckoned with.
Used appropriately, an LLM can behave like Sesame Street‘s sharper cousin. A student works with an idea and discovers something new. But it can just as easily behave like a vending machine that just talks back. The AI responds in a cold and transactional way. From across the room, these look the same. A child, a screen, some typing. Only one of them involves thinking, and you can’t always tell which from the outside.
This is the distinction I believe the screen-time debate keeps missing. Task-switching and saturation are real problems, but they’re not the whole story—maybe not even the main one. A student flipping through five tabs of polished answers and a student working with one hard question are technically doing the same activity. Both have AI open in a window. But I don’t think that’s even close to a meaningful similarity.
Friction Is the Curriculum
Learning is messy.
A student meets a new concept, misreads it, revises, objects to it, and even has to start over. The process is slow and maybe even a little embarrassing. (Which is probably why we keep trying to engineer it away.) But that struggle or friction isn’t an inefficiency in learning. It’s learning itself.
AI can support that process. It can also short-circuit it, and the short-circuit is seductive precisely because it sounds articulate and feels authoritative. The danger is that polished output reads like understanding even when no understanding occurred.
So maybe the line isn’t AI versus no AI. It’s whether the student supplied the “intellectual momentum” before the LLM showed up, or if the model supplied it for them.
A Rule of Thumb, Not a Policy
I don’t think this resolves into a clean classroom protocol and I’m not the teacher to suggest it. Further, I’d be suspicious of anyone who claims it does. But here’s a rough version. Have the student work the problem alone first, badly if necessary, for some time. Let them engage with AI with an actual position, even a wrong one. Then let the conversation do what conversations do, complicate things.
Afterward, ask the student to explain the idea with the screen closed. Ask where the model changed her mind, and if she can defend why it should have. For me, this ability to explain your work—from school to business—is the key aspect here that reveals the process of learning and not just the recitation of facts.
None of this is enforceable at scale, and I suspect most classrooms won’t bother because the school day wasn’t built for this slow pace. That’s a separate problem from the one I’m describing here, though probably not an unrelated one.
Beyond the Screen-Time Number
Parents already sense this distinction intuitively, even if they’d describe it differently. Not all screen activities feel the same, because they aren’t asking the same thing of the user.
Educational policy keeps measuring the wrong variable. Minutes are easy to count. Cognitive demand isn’t, which is probably the actual reason we keep counting minutes.
The harder question, the one I don’t think Horvath’s testimony or most of the coverage around it quite reaches, is what happens when the screen stops being a passive thing a child does something to, and becomes something that can do the thinking back.
Done well, that’s not a threat. It’s iterative engagement, the same mechanism that made Sesame Street work, except now the conversation never has to stop. A model that keeps pushing, keeps asking what you mean, keeps refusing to let a half-formed idea pass as a finished one, might be the most demanding and powerful tutor a student has ever had access to. And that’s very much the version worth protecting.
But it’s also the version most likely to get buried under the easier one. Iterative engagement takes longer than an answer. It asks more of the child, not less, and asking more is never the feature that gets marketed. We don’t have a Sesame Street study for whichever version actually wins out at scale. We’re running that experiment live, on our own children, and I’m not sure which version most of them are getting right now.
