This summer I hid something inside two of my assignment prompts.

I teach a couple of courses a year as an adjunct instructor, and I’d gotten tired of arguing about AI in the abstract. So I layered invisible instructions into the prompts themselves. Read the assignment the ordinary way, with your own eyes, and you’d never see them. Complete it that way and you’d turn in exactly the work I was hoping to grade. But copy that prompt into a chatbot, and the model would receive my hidden instruction alongside the visible one, and obediently follow both.

One of the hidden lines was deliberately mild. It changed nothing material and I marked it optional. One student noticed it and emailed to ask whether it was a real requirement, which is exactly the instinct I want a student to have. The second hidden line wasn’t mild. It would have redirected an entire project.

Eleven students submitted. Six of them handed back work responding to an assignment they couldn’t possibly have seen with the naked eye.

Graded on substance, they’d done the work. They hit the learning objectives. Several had clearly reworked the output before submitting it. Not one of them disclosed.

The easy read of that story is a scolding. That’s not what interests me.

What interests me is that we’re in the middle of a very loud national argument that has almost nothing to do with what happened in my classroom. Purdue now requires an AI working competency to graduate (Purdue University, 2025). Singapore’s national university made it compulsory for freshmen and handed the whole campus a chatbot. In the other direction, more than 260 organizations and experts signed an April letter calling for a five-year pause on all student-facing generative AI in PreK-12 schools (Fairplay, 2026). New York City council members pushed for a two-year pause of their own, and notably, they never argued the technology was bad for learning. They argued the city hadn’t protected student information in its contracts with AI companies (K-12 Dive, 2026).

All of that inside about ten weeks.

Read those fights closely and you notice they’re about two things: whether people may use it, and who gets the data. Very little of it is about honesty. Almost none of it is about what the tool does to the person using it.

is an educational neuroscientist who came on the show this week, and he draws the line in a place I find far more useful. When an expert hands a task to a machine, that’s offloading, because their expertise lets them evaluate whatever comes back. When a novice does the identical thing, that’s outsourcing, because nothing comes back that they’re equipped to judge. The tool doesn’t change. What changes is whether the person on the other end can tell when it’s wrong.

He tells it through his sister-in-law. She’s a gym rat. Her husband is not. They rented e-bikes to tour their city and the batteries died. In her case it didn’t matter. She was fit, so she pedaled the thing home. When his battery died, he called an Uber. He had no skill underneath the tool, so when the tool collapsed, he had no recourse. His new book, The Digital Delusion, came out this week, and it’s worth your time.

Same machine. Two outcomes, decided entirely by what the rider already had in their legs.

Which is why I keep getting stuck on the policing problem. We’re good at drawing measurable lines. We draw one at 0.08 for blood alcohol. At sixteen for a license. At eighteen for a lottery ticket. You clear those or you don’t. AI has soaked into so many ordinary tasks that no equivalent line exists.

The detectors don’t solve it. I’ve run writing through them and gotten back one hundred percent human on text I know a machine produced, and one hundred percent AI on text I know a person wrote by hand. Anecdotally, it feels wrong more often than right.

The newest attempt is a watermark. Anthropic announced this month that Claude will weave a statistical mark into its output as it generates (Anthropic, 2026). It’s real, and narrower than it sounds. It applies only to models released on or after August 2, no detection tool has shipped, and the company says plainly that a mark shows the model processed the text, not that it wrote it.

That last part is where this becomes a double-edged sword. I disagree my verb tenses constantly. I slide between you and I and we and one. You can probably see it in this piece. If I write something myself and hand it to a machine to fix the grammar, most people would call that fine, because that’s what a human editor does. A watermark would still say a machine touched it.

Here’s where I’ve landed, personally. I’m writing my next book in a physical, college-ruled notebook. I’m minding the margin, that faint magenta line my fourth grade teacher penalized me for crossing, double spacing so I can go back and fix things in between. I stopped using Grammarly, because it isn’t telling me I confused my verb tenses anymore. It’s rewriting my sentences. It’s replacing the humanness, however flawed, in me.

So where’s the line?

I’m not sure I have a clean answer. But I’m fairly sure we’re arguing about the wrong thing when we argue about how much is too much, and the question sitting underneath it is a harder one.

I worked the rest of this out in the full essay, including the student who looked me in the eye over Zoom and asked me what AI was. You can read it here.

If this is the kind of question you like sitting with, Collide goes out every week. And The Ryan Vet Show is where I ask people like Jared these same questions out loud.

A five-year-old I know can say “hey Siri” before she can spell it. She asks the machine on the kitchen counter how far away school is, whether it will rain, what sound a narwhal makes. She has never once needed to be read to in order to find something out. She just asks, and then keeps asking, because the device will hold the conversation as long as she wants.

That small scene is playing out in kitchens all over the country this month, because a new class is starting kindergarten. Most of these children were born in 2021. They are the real COVID babies, the ones conceived during the pandemic rather than simply born beside it. And they are beginning a thirteen-year journey that will end, if the calendar holds, with the graduating class of 2039.

Here is the detail that reframes the whole thing. 74.5% of 2021 births went to a Millennial mother (Osterman et al., 2023). This is the first class raised almost entirely by Millennial parents, the generation that carried Razr flip phones through high school and still remembers the exact sound of dial-up. They are now raising children who will never know a world without a machine that talks back.

I find it useful, every so often, to stop looking at the graduates and look instead at the kids walking in the front door. What they are walking into tells you more about the next twenty years than any trend report.

Start with what is being taken away from them, on purpose. Three years ago, a single state restricted student phones in the classroom. As of this year, 38 states plus DC require districts to ban or restrict them, and most of those go bell to bell (Dalbey, 2026; Rock, 2026). Their parents carried a phone in their pocket through high school. These kids may go twelve years without one in a classroom. After a decade of blaming devices for what happened to young attention spans, this is the first cohort raised inside the correction.

Now set that against what is being handed to them, just as fast. By early 2026, most state education departments had issued formal guidance on artificial intelligence in schools, guidance that did not exist in 2022. The question educators ask is no longer whether these children use AI. It is when.

So one hand is pulling friction back into childhood while the other removes it at a speed no generation has ever experienced. Less distraction. Less struggle. At the same time. I keep turning that contradiction over.

There are smaller signals too, and they stack up quickly. The homeschooled share of K-12 students doubled between 1999 and 2023, from 1.7% to 3.4% (Bielick et al., 2001; Vaden-Kiernan et al., 2024). Add charters, microschools, hybrids, and virtual options, and this class will not graduate from one system. It will graduate from a dozen versions of school, most of which barely existed when their parents sat in those seats.

Even driving is on the table. North Carolina separated the driver from the license back in 2017, before a single child in this class was born (N.C. Gen. Stat. 20-401). By the time these kindergartners reach sixteen, there is a real chance driver’s ed is optional, because the car drives itself.

None of these on their own is dramatic. Together they describe a childhood with a fundamentally different physics than the one their parents knew. Newton said every action has an equal and opposite reaction. I have come to think that holds for generations, not just objects. Every convenience we hand a child quietly removes a struggle that used to shape them.

I laid all eight out in this week’s Collide essay: https://ryanvet.com/collide/millennials-are-raising-the-class-of-2039/

But there is one I have not been able to shake. It is the eighth, the thing this generation can do that no generation before it could, and what it means for a five-year-old standing on a playground.

That is where I want to stop, because it deserves the full piece and not a paraphrase.

If you think about generations, culture, and where all of this is heading, subscribe to Collide. One essay a week, calm and research-grounded, no alarms. I would love to have you along.

There’s a stretch of highway near my house where the police recently put up temporary cameras on twenty-foot poles. White box, solar panel, a little blue light on top. Someone had staked a piece of corrugated plastic in the grass in front of them and written across it in Sharpie: Surveillance is bad.

It’d be easy to file that under paranoia. Somebody annoyed that their morning commute is getting logged. But the more time I spend with the research on being watched, the more I think that crude yard sign sits closer to the evidence than most of the confident arguments for watching everything.

We tend to treat surveillance as neutral at worst and protective at best. Watch more, and you get more safety, more accountability, more truth. The studies keep finding something closer to the opposite.

Start with how early the instinct to perform shows up. Children become sensitive to being evaluated by other people by twenty-four months (Botto and Rochat, 2018). In one experiment, a five-year-old left alone in a room took a sticker that belonged to another kid about a quarter of the time. Put a second child in the room, and that dropped to four percent. The researchers concluded that by age five, kids are already actively managing their reputations (Engelmann et al., 2012). The urge to behave differently when someone’s watching is installed before most of them can read a full sentence.

Now look at what we’ve built around this generation. Cameras in the playroom that announce, by face, who just walked in. Watches that are also phones and also location trackers strapped to a wrist. Shoes and Crocs, yes, Crocs, designed with a slot for an AirTag. A kid today can be found, as a dot on a map, at any moment of any day. We call it safety, and we mean it sincerely.

Here’s the uncomfortable part. We’ve never seriously studied what an entire childhood under observation does to a person. The closest thing we’ve got to a natural experiment is child stardom, growing up with the eyes always on. Mara Wilson, who played the girl in Matilda, once described that life as having to live up to a fan base, “a little like having to deal with a million strict parents who don’t actually love you” (Wilson, 2013). She was describing the feeling of being watched even when the room was empty. The audience never really leaves.

The workplace runs a quieter version of the same experiment on adults. A researcher spent eighteen months inside one of the largest phone factories in the world and found that workers hid their best, fastest methods the instant a manager appeared, then switched back to the slower official script. One of them asked for curtains around the line, not to conceal anything, just to feel unwatched for a while. Defect-free output rose ten to fifteen percent and held that lead for five months (Bernstein, 2012). Performance improved when people stopped feeling watched. A separate team found monitored employees were more likely, not less, to break the rules, cut corners, and slow down on purpose (Thiel et al., 2023). Strip people of the sense that they author their own conduct, and they’ll hand it to whoever’s holding the camera.

There’s a name for what all of this quietly costs. Patrick Lencioni splits trust in two. Predictive trust is knowing someone will do what they said they’d do. Vulnerability-based trust is someone choosing to tell you the hard thing, the mistake, the fear, before you ever find it yourself. The camera nails the first kind. The location dot proves she was where she said. The doorbell proves the package came. The call recording proves the script got followed. And in delivering all that proof, the machine quietly removes the moment a person decides, on their own, to be honest with you.

Lencioni also lists what a team without that second kind of trust looks like. People hide their mistakes. They hesitate to ask for help. They jump to conclusions about each other’s intentions. That last one is the quiet danger of a monitored world. The machine hands us a fact, stripped of context, and we fill in the story ourselves, usually without ever asking the person what actually happened.

You’re left with the truth and none of the trust.

So what do we actually do with tools that aren’t going anywhere, and what does a childhood spent on a leash cost the generation living it? That turn, the part about my daughter learning to climb the stairs and what changes the moment we look away, is where the full essay lands, at ryanvet.com/collide/gen-alpha-is-the-experiment/

If you want one idea, researched properly, every week, subscribe here and consider joining the Collide newsletter at ryanvet.com/collide.

It took Rudy Kurniawan the better part of a decade to become the most successful wine counterfeiter in history. Known in collecting circles as Dr. Conti, he blended cheap wine in his California kitch…

Read more

Read more

There is a phrase making the rounds in my house lately, and it stopped me cold the first time I heard it.

“Just check the camera, Dad.”

It came from one of my sons, mid-argument with his brother, over …

Read more

For almost thirty minutes on the drive home, I interrogated my six-year-old about a goat.

Had he painted a picture of a goat, or painted an actual goat? With what? What colors? Why? What did the goat …

Read more

If you are like me, when you imagine the signing of the Declaration of Independence, you picture John Trumbull’s painting. The one on the back of the two dollar bill. The calm and orderly men in powd…

Read more

“AI slop,” the comment read, with a laughing emoji. It was the second nearly identical one, and both came from accounts that looked like bots. The irony was hard to miss. A bot was accusing me of bei…

Read more

The men who wrote the wisest fathers in television history had survived a depression and a war. Ed James, Joe Connelly, Bob Mosher, Sherwood Schwartz. Most of them were born between 1907 and 1917. Th…

Read more

In August of 1939, audiences sat in dim theaters and watched a gray Kansas farm fill the screen. A girl, a dog, a coming storm. Then the tornado lifts the house, sets it down somewhere new, and Dorot…

Read more

Read more

That meant, if we did happen to turn on the TV, we usually watched PBS.

Bob Ross. Sesame Street. The occasional documentary I was probably too young to appreciate. And…

Read more

If George Orwell were alive today, he wouldn’t need to imagine 1984. He would just open TikTok.

He would see the cameras watching our doors, our schools, our streets. And he would notice something he …

Read more

If the average graduate is 22, then most were born in 2004. That number is a useful key.

Read more

A teenager in …

Read more