QuitGPT' protesters rally outside OpenAI HQ in San Francisco over deal with  Pentagon - Local News Matters

There's a survey going around that supposedly proves my generation is souring on AI. It's the new Gallup study with the Walton Family Foundation and GSV, more than 1,500 of us between 14 and 29, and the headline number is that weekly AI use among Gen Z has basically flatlined. Just over half of us use it weekly or more, which is almost exactly where it was a year ago. Excitement dropped 14 points. Hopefulness fell nine. Anger went up nine. Nearly half of working Gen Zers now say the risks of AI outweigh the benefits, up from 37% a year ago. The report calls it "the AI paradox," and most of the coverage I've seen rounds it off to a clean story: the generation that grew up with this stuff is getting cold feet.

I sit in classrooms with these people every day, and I think that read is wrong, and not just a little wrong.

The flat average is hiding a split. What I see at school is two groups pulling apart from each other, and when you average them you get a number that describes neither one. That doesn’t mean that there hasn’t been turnover and change since last year as these tools and habits begin to diffuse further; there has been.

The two kids

Let me describe two people I actually know. I’ll change enough that they stay anonymous, but the archetype is accurate and you probably know versions of them too.

The first one, call him Marcus, runs everything through ChatGPT. Essays, lab write-ups, the email to the coach, the cover letter for the summer internship. He is fast. His output volume is genuinely higher than mine. And about a month ago he got handed back a history essay with a comment from the teacher asking him to explain his second paragraph out loud, and he couldn’t. He had written it, submitted it, gotten a B-plus, and he could not reconstruct the argument because he had never really held it in his head. He laughed it off. “I wrote it, I just don’t remember it,” he said.

The second one, call her Mary, uses AI just as much. Maybe more. But she treats every output like a first draft from an intern she doesn’t fully trust. She’ll generate four versions of an opening, throw out three, rewrite the fourth by hand. She uses it to find the counterargument she didn’t think of, then goes and checks whether the counterargument is real. When she gets something back, the first thing she does is look for the part that’s wrong, because she’s learned there’s almost always a part that’s wrong. Her grades are better than Marcus’s. More importantly, she can still explain her own second paragraph.

Marcus and Priya use AI roughly the same amount. On a survey they’d both check “daily.” But one of them is getting measurably sharper and one of them is quietly hollowing out, and the volume number can’t tell you which is which.

Why the averages mislead

This is the thing about the Gallup data that I think people are skipping past. The negative sentiment isn’t randomly spread. It tracks how people use the tool. Among daily users, 69% feel curious, 44% excited, 38% hopeful. Among people who never touch it, those numbers are 28%, 4%, and 2%. The anxiety gap is even bigger: 60% of non-users feel anxious about AI versus 28% of daily users. The people closest to it are the least afraid of it, which is not what you’d expect if proximity bred contempt.

So when the report says excitement dropped 14 points across the board, that’s a real finding, but it’s an aggregate of a population that is internally splitting. Some of that drop is people like Mary getting more clear-eyed and less starry, which is healthy. Some of it is people like Marcus sensing, correctly, that something is off about how they’re using it. And a lot of it is the 19% who never use AI at all getting louder and angrier as the tool becomes unavoidable.

When you blend the person who’s learning to wield it, the person who’s being eroded by it, and the person refusing to touch it, you get a flat, gloomy, contradictory average. That average is what the report calls the paradox. What’s actually underneath it is three different populations getting mashed into one number that describes none of them very well.

The actual thesis

Here is the claim I’m going to spend the next two essays defending, so I want to state it plainly now.

More volume is genuinely better. The instinct to cap your AI use, to ration it on principle because you’re worried it’ll make you lazy, is mostly a mistake. The students I know who are pulling ahead are not the ones using AI less. They’re using it constantly. Volume gives you more shots, more drafts, more angles, more things to react to.

But volume only compounds in your favor if you keep one thing in your own hands, which is the judgment about where the human touch has to go back in. Where to override the model. Where it’s confidently wrong. Where the “good enough” draft is actually a C-minus draft dressed up in clean sentences. That layer of taste and verification and knowing-where-to-look is the whole game. People who have it can run unlimited volume and get stronger. People who don’t have it run the same volume and get weaker, because every task they hand over completely is a task they no longer know how to do.

So the question isn’t how much AI you use. Marcus and Mary answer that one identically. The question is whether you’re building the judgment layer or letting it atrophy, and that’s a question almost nobody is measuring, which is why the surveys keep producing numbers that feel true and explain nothing.

What this means if you use AI

The practical takeaway from this first piece is small but it reframes everything that follows. Stop asking yourself whether you use AI too much because that’s the wrong dial. Ask whether you can still do, unassisted, the core thing the AI is doing for you. If Marcus had to write a history essay with his laptop closed, could he? If the answer is no, you’ve found something worth fixing, and your AI use is what made it visible. Use the tool as much as you want, and separately, protect a set of reps where you keep your own hands on the work. I’ll get specific about how in part two.

What this means if you’re building or investing

If you run a business and you’re staring at this same Gen Z data trying to decide whether the AI consumer wave is real, the flat adoption number is going to tempt you into one of two dumb conclusions. Either “young people are cooling on AI, the hype is over,” or “they’ll come around, just keep shipping features.” Both miss the split.

The real signal is that your most engaged young users are bifurcating into people who extract enormous value from your product and people who are passively consuming it and slowly resenting it. Those two cohorts will have wildly different retention, willingness to pay, and word of mouth, and your top-line “weekly active” number averages them into mush, exactly like Gallup’s did. The companies that win the next few years are going to be the ones that figure out how to build the judgment layer into the product itself, so that using more of it makes the user better rather than more dependent. That is a specific, buildable thing, and it’s where I think the durable money is. I’ll make the full investment case in part three.

For now, the one move worth making is to stop reading flat engagement as flat value. Under that flat line, your users are sorting themselves into your best customers and your future churn, and right now almost nobody is instrumenting for the difference.

Next, in part two: the mechanism underneath all of this. Why the kid who can’t quote his own essay is a preview of a real and measurable problem, what the EEG studies actually found, and what the “human fine touch” is made of when you look closely. After that, in part three, where I’d put my money if judgment is the scarce asset.