Tuesday night I used AI to build a Dungeons & Dragons character. I told it what I wanted and it handed me back a rogue with a tragic childhood and a chip on her shoulder, fully statted, in about four minutes. I felt a little guilty. Then I felt guilty about feeling guilty. That's about the most 2026 sequence of emotions a person can run through before dinner.

Here's the thing about that guilt: I teach people for a living. Not kids, not in a classroom with a bell — grown adults in corporate rooms who are getting paid to be there. So I've been watching the panic about AI and education with more than a casual interest, because the thing everyone's terrified AI is stripping out of students — the ability to think — is the exact thing I spend all day dragging out of grown adults. Challenging how people think is the whole job. And I go hardest at the ones who've stopped doing it.

And the headlines this month are a whole vibe. Students running AI agents that log into Canvas and "take" the entire online class — watch the lectures, write the papers, even pose as the student in the discussion boards — while the actual human doomscrolls in another tab. Teachers warning that kids are losing the ability to think because they've outsourced the struggle of writing — and writing, as Dana Goldstein put it in the New York Times, quoting cognitive psychologist Ronald Kellogg, is a technology for thinking. Wring the labor out of it and you're not writing anymore, and by the same logic, not thinking.

I'm not going to tell you that's wrong. It isn't. There's an MIT study floating around showing people who wrote with AI had lower brain activity and couldn't remember a single line of what they'd just "written." I believe it. If you never dig the ditch, your back never gets strong. Fine.

But you'll forgive me if I've heard this song before. We've been clutching these same pearls for twenty years, at least. The internet was going to rot our memory. Before that it was the calculator — kids won't be able to do math in their heads. Before that, somebody somewhere was sure the slide rule would make us soft. They don't even teach long division anymore, because what, exactly, is the point? Every generation discovers a new machine, decides it's the one that finally hollows out the human mind, and writes the same worried essay about it. The essay is always well-argued. The mind keeps working anyway.

So when I say I used AI to build a D&D character, and my husband is over there building an entire online course on electronic music for his nephew — leaning hard on Claude to do it — and neither one of us feels like we're committing a crime against thought, I know I'm being a little facetious. Because there is a difference between what he's doing and a kid renting a bot to sit through Sociology 101 for him. I know there's a difference. (Can you see my deadpan stare? Of course I know there's a difference.) I just think we've been pointing at the wrong one.

The difference was never the tool. It's what you're doing with it.

Two things, really. The first is honesty. And the sharpest statement of it this month didn't come from an educator — it came from a judge. A man representing himself in a Connecticut court hid instructions in a legal filing, tiny 3-point white text buried in the document, telling any AI that read it to side with him. (404 Media's Jason Koebler caught the story; a human clerk caught the injection, having noticed some odd white space.) The judge was not impressed. But read what he actually objected to: Used honestly, he wrote, these tools hold real promise — a person who can't afford a lawyer can now put a coherent argument in front of a court. What this man did was use the tool dishonestly: he transmitted a hidden second message engineered to change how his filing was judged. The problem wasn't the AI. The problem was the lie.

That's the whole line, right there. Compare it to the expert witness who got obliterated when the opposing team subpoenaed his ChatGPT logs (reported, again, by 404 Media) — a guy with twenty years of supposed experience who had the machine write his entire report defending a corporation, then uploaded a photo of the device at the center of the case and asked ChatGPT, "what am I looking at?" He didn't use a tool to think harder. He used it to not be in the room at all, and then he lied about being there. That's not a thinking problem. That's a character problem wearing a thinking problem's coat.

The second difference is one nobody clutching pearls wants to talk about, because it cuts the other way: sometimes the machine is where the thinking starts.

I want to go back to that man and his 3-point white font for a second, because he's disgraceful and also, I have to admit, kind of a genius. Hiding a prompt injection inside a court document so that any AI reading it gets quietly told to rule your way — that is not the work of a person who has stopped thinking. That is Machiavelli with a word processor. It's dishonest, it's beneath the dignity of the court, and it is, flatly, creative. Crime and porn really do lead the way, don't they? Every technology's most inventive early users are the people trying to get away with something. That's not an argument for the crime. It's a reminder that the same tool the panic says is killing our creativity is the one people are using to invent things nobody thought of yesterday — including, God help us, new ways to cheat.

And here is where I stop being nervous about all of it. Because for all the noise about the machine taking over our minds, the machine is a spectacularly bad learner. MIT Technology Review talked to Michael C. Frank, a cognitive scientist at Stanford, about how these models actually acquire language, and the number is humbling. (That piece is paywalled; Futurism wrote it up for free.) A toddler starts producing real sentences after soaking up something like ten to thirty million words — picked up in a living room, over the course of a year, between naps. To fake that same milestone, Frank says, we have to "burn down a forest and scrape the entire sum of all human knowledge." Train the model on thirty million words, the amount a one-year-old runs on, and you don't get a kid. You get a nonsense generator.

Sit with that. The thing we're terrified is going to eat our children's ability to think cannot do what a one-year-old does before it can walk. It is not a superior mind. It is a very expensive, very confident mirror with a forest fire behind it.

So no, I don't lie awake worried the machine is going to think for us. It can't. What I worry about is the reflex underneath the whole panic — the one that shows up loudest in the stories about homeschool parents generating an entire curriculum from a chatbot to match their exact worldview - no history they don't like, no science they don't believe. That's real, and it's grim. But the reflex isn't new either. It's the oldest one there is: I know best, and different means wrong. The machine didn't invent that. It just gave it a faster printer. You can see the same reflex running our politics right now, and I'd argue it's done a lot more damage there than any kid with a chatbot ever will.

I don't have children. I have a niece I would walk through fire for. And if you put a tool in front of her that helped her think — differently than I think, in a shape I don't recognize, arriving at answers I wouldn't have reached — I'd be the first person to fight for her access. Because she doesn't think exactly the way I do, and the day I decide that makes her wrong is the day I've stopped being any use to her.

The question was never whether the machine can think for you. Ask a toddler. It can't. The only question that was ever worth asking is whether you still are thinking for yourself — honestly, in your own name, with your own hands somewhere on the work.

So. What's the difference?

You already know. It's whether you're still in the room.