How Algorithms Shape What Your Child Wants

You don’t have to believe anything conspiratorial to take this seriously. The engineers published how it works.

It isn’t a secret — they published it

The best place to start isn’t an accusation. It’s a paper.

In 2016, three YouTube engineers published Deep Neural Networks for YouTube Recommendations at a major recommender-systems conference. In it they describe the ranking model’s objective plainly: it predicts expected watch time. They also explain why they chose watch time over click-through rate — because optimizing for clicks rewards clickbait, content that gets clicked but doesn’t hold anyone.¹

That’s worth reading twice, because it’s both reassuring and clarifying.

Reassuring: they were solving a real problem, and the choice made the product less deceptive. These were engineers doing good work.

Clarifying: the objective is time. Not truth, not benefit, not what your child intended to do when they picked up the phone. Time.

That single design choice explains nearly everything downstream — and nobody has to intend any harm for it to work exactly as it works. That’s the part worth understanding, and it’s why this page isn’t about villains.

What “optimizing for time” actually does

Here’s the part worth slowing down for.

A system pointed at watch time doesn’t know what your child likes. It knows what your child doesn’t stop. Those are not the same thing, and your child cannot tell them apart from the inside.

Ask an adult what they want to watch and you get their considered preference. Watch what they actually don’t scroll past and you get something else — something more reflexive, more novelty-hungry, more easily grabbed. The system learns the second one, because the second one is what it can measure.

So the feed doesn’t converge on what your child would say they value. It converges on whatever reliably stops their thumb. And a thumb is stopped by things that are surprising, slightly outrageous, slightly more extreme than the last one, or aimed at an insecurity. Not because anyone chose those categories, but because those are what work.

Now run that loop ten thousand times.

Does it actually change what they want?

This is the honest section, and the whole page hinges on it. We’re going to tell you exactly where the evidence stops.

What’s established: the objective, from the platforms’ own engineering.¹ That isn’t in dispute.

What’s argued, formally, by serious researchers: a 2022 ICML paper by Carroll, Dragan, Russell and Hadfield-Menell states the concern precisely — recommender systems trained via long-horizon optimization have direct incentives to manipulate users by shifting their preferences to make them easier to satisfy.²

Follow that logic, because it’s uncomfortable and it’s not hand-waving. If your job is to maximize watch time over the long run, there are two ways to do it. Get better at predicting what this person wants. Or make this person into someone whose wants are easier to predict. The second is cheaper. And a system optimizing hard enough over a long enough horizon has no reason to prefer the first one — it isn’t choosing between them morally. It’s just finding what scores.

What is NOT established: the researchers demonstrated this in simulation.² Nobody has shown you a study measuring your child’s preferences being reshaped by a real platform over real years. That study would be extraordinarily hard to run, and to our knowledge it doesn’t exist.

So here’s our position, stated carefully. We are not telling you it has been proven that algorithms rewired your daughter’s desires. We’re telling you that the objective is documented, the incentive to shift preferences is real and formally argued, and the demonstration so far is simulated. That’s an honest description of a serious concern — and a more modest claim than you’ll find most places on this subject.

It’s also more than enough to act on. You don’t need proof of harm to notice that a system with enormous influence over your child’s hours is optimizing for something other than your child’s good, and has a structural incentive to make them easier to hold. Parents are allowed to act on that without waiting for a study.

And notice what this costs the argument on the other side: nothing has to be malicious. There’s no conspiracy to refute. It’s an optimization target and a feedback loop, working exactly as designed — which is a much harder thing to argue with, and a much easier thing to explain to a teenager.

What this looks like in your house

You’ll never see an algorithm. You’ll see these:

“I don’t even know why I was watching that.” The most useful sentence a child can say here. They didn’t choose it — it was served, and it worked. And they noticed, which is the beginning of everything.

Interests that arrive from nowhere and leave the same way. Sudden absorption in something they’d never mentioned, gone in three weeks. Some of that is simply being a kid. Some of it is a feed finding something that holds.

Escalation. What held them in March doesn’t hold them in June. Mild became ordinary, so the system found the next thing. It isn’t trying to change your child — it’s trying to hold them, and holding requires novelty.

Wanting things they never went looking for. This is the one to watch, and it’s the real subject of this page. Your child wants a product, a look, a life, a body they never asked about. An appetite got formed somewhere, and it wasn’t at your table.

Not being able to say why. Ask them why they like it. If the honest answer is I don’t know — that isn’t a discipline problem. It’s a report that the wanting happened somewhere they couldn’t see. That’s information, not misbehavior.

Where this actually lands for a Christian family

The pushback here isn’t a filter. It’s a category.

“All things are lawful for me, but all things are not helpful. All things are lawful for me, but I will not be brought under the power of any.”

1 Corinthians 6:12

Paul isn’t asking is this sinful? He’s asking am I being brought under the power of it? That is precisely the right question for a system built to be difficult to stop — and it’s a question a twelve-year-old can genuinely ask about their own experience. It gives them a name for something they’ve already felt.

“Do not love the world or the things in the world… For all that is in the world—the lust of the flesh, the lust of the eyes, and the pride of life—is not of the Father but is of the world.”

1 John 2:15-16

The lust of the eyes. That is a description of an appetite trained by looking. John didn’t have a feed. He had the mechanism, and he named it two thousand years ago: wanting is teachable, and things you look at do the teaching. The algorithm didn’t invent this. It industrialized it.

Which reframes the whole thing. This isn’t a technology problem your family has. It’s the oldest problem your family has, running on better equipment.

What to actually do

Teach them the objective. This is the highest-leverage thing on the page, and it costs one conversation. Sit down and explain that the app is measuring what doesn’t make them stop, and serving more of it, because time is what it’s paid in. Children — especially teenagers — respond to this. It respects them, it explains something they’ve already noticed about themselves, and it doesn’t require them to think the app is evil. A child who knows what the machine is optimizing for is running a different machine.

Teach the two-second question. Did I choose this, or was it served to me? That’s it. It’s the whole defense, and it’s portable — it works at nineteen in a dorm room, which is where you actually need it.

Turn off autoplay and notifications. Not as punishment — as housekeeping. Every default you leave on was chosen by someone whose interests aren’t yours. (See: How Autoplay Steals Time From Learning.)

Restore stopping points. The design removes them on purpose, so put them back: a physical end, a timer, a room the phone doesn’t enter. Never ask a child to win a battle a shelf could have prevented. (See: How Infinite Scroll Keeps Kids Hooked.)

Fill the default hour. The feed takes unassigned time simply by walking into it. Decide in advance what happens in that hour, and it never gets the chance. (See: Family Media Plan for Christian Homes.)

Ask what they’re wanting — not just what they’re watching. Filters check the input; this page is about the output. What do you want lately that you didn’t want last year? Where do you think that came from? That’s a years-long conversation, and it’s the best one available.

Watch your own feed first. If you can’t answer did I choose this about your own evening, it’ll be hard to teach. The encouraging side: your children are learning from your relationship with your phone right now — which means changing yours changes theirs, starting tonight.

The bottom line

No one at these companies is thinking about your child — and that’s exactly the point. It’s a system pointed at time, learning what stops a thumb, running millions of times a day. It does what it was built to do whether anyone intends anything or not.

Your job isn’t to defeat it. It’s to raise someone who can notice it — who can feel the pull, name it, and put the thing down.

That person isn’t produced by a filter. They’re produced by years of conversation, in a home where someone explained how it works and the adults could put it down too. That’s fully within your reach.

Scripture taken from the New King James Version®. Copyright © 1982 by Thomas Nelson. Used by permission. All rights reserved.