The Hours Were Never the Problem

The quantity debate outlived its own evidence by ten years. What matters is not how many hours, but what a tool is built to maximize.

For twenty years the argument about children and screens was an argument about a number. Two hours. Then some other number, adjusted for age, for homework, for whatever a household could enforce on a school night. Parents kept score. Schools wrote it into handbooks. Kids chafed against it.

The rule stopped being the pediatric advice in 2016. Ten years ago. The American Academy of Pediatrics has said since, in one document after another, that a rule about hours does not cover what children and teens actually need in order to have a healthy relationship with media. Its current parent guidance puts five questions where the number used to be. Who is this particular child. What is the content. Has media become the only way this kid calms down. What is it crowding out. Is anyone talking about it.

The household argument went on anyway. The quantity debate did not end when the guidance moved; it just stopped having anything behind it. And in January the academy went further, announcing a new policy statement that puts the design of the product, rather than the duration of the visit, at the center of the problem. That should have been news. It mostly was not.

Why the number was so easy to argue about

Because it was measurable, and it let everyone win something. A parent could comply with two hours. A school could enforce it. A skeptic could point out that two hours of reading and two hours of autoplay are not the same thing, be obviously right, and change nothing.

It also gave us a fight we could have entirely at home. Count the hours, set the timer, take the tablet at nine. Every one of those moves is aimed at the child. Not one of them is aimed at the product.

What a time limit actually puts on trial

A screen time limit is a verdict on a kid's behavior. It says the child used too much. It does not ask what the thing was built to do to the child while it had them.

The academy is more direct about this than the debate ever was. Its own announcement says most platforms are designed to boost engagement and profit, not to support children's health and development, and the pediatrician who led the policy puts the mechanism plainly: many platforms are guided by an underlying business model to keep users engaged for as long as possible. The design follows the model. Autoplay. The next episode already loading. The feed with no bottom. The streak that punishes a missed day. The AAP now asks manufacturers to turn autoplay off by default, which tells you it is a design decision and not a fact of nature. None of that is a parenting failure and none of it is measured by a clock. A family can be perfectly compliant with any hour limit you name and still hand a child four hours a week of something engineered to be difficult to leave.

I am not standing outside this. I helped raise three kids while working too much, and I know what autoplay was worth on a bad evening. I was not watching the design, the clock, or the child.

Then the promise arrived

The same design logic has now moved into something sold as the opposite of a time sink. AI arrives in a child's life wrapped in the language of learning. It will tutor. It will explain. It will meet a student where they are.

I have reviewed dozens of these products professionally. Some of them are going to do real good for kids, and I want to say that first, because those need to be celebrated and not written off. Others are built to do something else entirely, which is to assemble a marketing profile on a child and then sell to it. The pattern is consistent enough to describe. The marketing claims things the product does not do. The agreement runs long and deliberately hard to read in exactly the places where one plain sentence would tell a parent what is being collected and who gets it. And the trust signals do not survive a phone call. I have seen a landing page claim a partnership with one of the largest school districts in the country. I have seen another claim a national youth organization that any parent would recognize on sight. Those badges do work that no product feature could do, and almost nobody checks them.

A parent cannot realistically phone a children's charity to confirm a logo on a landing page, and I would not ask you to. The rest of what you need is sitting in the open. Open the user agreement before your kid opens the app. Ask the company in writing what it collects and who it sells to, and notice whether the answer is a sentence or a paragraph that dissolves when you read it twice. If a product is good for your child, it survives fifteen minutes of your attention.

Start with the badges, because they are the easiest thing to falsify. Most districts keep a list of approved vendors and partner relationships, and seeing it does not take a phone call. Email the purchasing office, or technology, or whatever that function is called in that district, and ask one question: does this company appear on your approved list.

Then read the answer correctly. A confirmation costs you ten minutes. A correction is worth a phone call. And if nobody there can find any record of the company at all, that is an answer from a person who has to own what they say. A partnership that leaves no trace inside the organization it names is a marketing asset, not a partnership. And it means something more specific than a false claim. Nobody at that organization has reviewed this product and decided it was safe or appropriate for a child. The badge does the exact work a review would do, and the review never happened.

And take none of it on trust, including from me. Not from a news segment, not from the store rating, not from the parent group where somebody's cousin swears by it. Any of us can be wrong about a product we liked, and you are the one who lives with the answer. Checking was never about being handed the right answer. It is the part that stays yours.

Look at what kids report doing with it. Common Sense Media surveyed 1,204 children age 9 to 17 in March. Eighty-six percent use AI, and about a quarter use it every day. Among the ones who use it, 89% use it for entertainment and 85% for help with schoolwork, and close to half of them are doing the schoolwork at least once a week. Both things at once, on the same product, with nothing in between them.

Two more findings from that survey belong next to those. Just 35% of the kids know AI cannot tell the difference between what is true and what is false, so the homework use is running through a population that mostly believes the tool cannot be wrong. And 44% say no parent or guardian has ever talked with them about how to use AI safely.

The engagement question is being asked, just not by us. Last September federal regulators sent orders to seven companies whose chatbots act as companions, asking how they monetize engagement, how they test and monitor for negative effects, and how they mitigate harms to children. That is a regulator asking what a product is optimized for. It is a better question than how many hours, and it is not the one being asked at the kitchen table.

Same hours, two different children

Here is the experiment that should have ended the quantity debate.

Researchers at Penn ran nearly a thousand high school math students through practice under three conditions: no AI, a plain chatbot, or a version built as a tutor, which gave hints instead of answers and had teacher input behind it. During practice, the plain chatbot group scored 48% above the students working alone. The tutor group scored 127% above them. Then the tools were taken away, and all three groups took the same exam.

The plain chatbot group scored 17% worse than the students who never had AI at all. The tutor group showed no such penalty.

Same subject. Same amount of time. Same kids, randomly assigned. The only variable was what the tool did when a student got stuck, and it separated the student who had practiced from the student who had watched. The work left the room with the tool.

That study is peer reviewed, which is worth saying because a single striking study is where this publication usually gets careful. The evidence is fairly clear on what to look at, and it is not the clock.

The debate worth having

Three questions replace counting hours.

What is this built to maximize. Everything has an objective, and if you cannot tell what it is after ten minutes of using it, assume it is attention.

Does it hand over the answer or the next step. That is the whole distance between the two groups in that experiment.

Does it get harder as the student gets better. A tool that stays equally easy forever is entertainment, whatever the marketing says.

The quantity debate outlived its own evidence by ten years because it gave us a number to argue about instead of a product to evaluate. The hours were never the problem. What the hours were building, or quietly failing to build, always was.

What the checking turned up

Three things about the evidence behind this piece, because you should be able to go look.

The Common Sense Media figures in this piece come from the report, not the press release, and the two do not read the same. The release says "only a third" of kids know AI cannot tell true from false. The report says 35%. The release says "well over 8 in 10" AI users use it for entertainment and homework help. The report says 89% and 85%. Nothing is wrong in either one. But a press release rounds, and rounding is not a ceiling on what the study found.

The Penn study is peer reviewed, and its correction notice turned out to be an author's institutional affiliation and nothing else. Worth knowing for anyone citing it: the published abstract says only "nearly a thousand high school math students." Details you may have seen elsewhere about where that school was are not in the published paper. This piece claims only what the abstract does.

A second trial out of Germany last December, which this piece does not cite, gets used as evidence against making students read before they touch an AI tutor. Its abstract says unrestricted access "significantly" outperformed restricted access. Its own results section says that difference is marginally significant, p equals 0.066. The abstract and the body do not agree, and the body is the paper.

When we get something wrong in a published issue, it will be listed here too.