Essay
A client sent me an eight-page foreword written by ChatGPT and needs four rounds for something he could have done in one. That same morning, another client proves the opposite. The tool is identical, the outcomes are opposed. People aren't getting stupider because of AI. They're getting lazy where judgement lives, and that's the most expensive place there is.
A client sends me the foreword to his style guide. Eight pages. I read the first paragraph. By paragraph four the same thought appears again, worded slightly differently, and by paragraph seven it's there for a third time. It's not a text, it's an echo listening to itself. And what foreword actually needs eight pages?
I email him back telling him to pack in the ChatGPT. He agrees. He writes that a day later, in an email where he immediately asks for a compromise. Then he sends a new version, this time written by himself. It's good.
Count along with me. One foreword. Four rounds. Two of them spent discussing why the first piece wasn't up to scratch.
That's not an opinion about artificial intelligence. That's an extra invoice.
In the printing trade there's a moment between making and printing. The proof. You get a sheet back, or an entire book. You lay a ruler along it. You actually look at it. You read the text through thoroughly one more time with a pencil and a marker in your hand. Only then does the press run a thousand sheets.
That proof isn't the product. Everyone knows that. The proof exists to show you what might still be wrong. Lots of people now send me the proof straight away, as though it were the print run.
A language model always delivers a perfect proof. A tidy, flowing, immediately readable proof, and that's precisely the problem. It looks finished. You recognise a half-baked sketch as a sketch immediately. Eight pages of impeccable Dutch you don't. There's not a single signal in the text that says: 'Someone still needs to look at this!'.
My client doesn't skip that step because he's lazy. He skips it because nothing reminds him there still is a step. His own judgement. And his own vision. He cheerfully chucks in the bin the things we used to value so highly until recently.
Because that's what that step is. Reading. Not scanning, not flicking through looking for something that jumps out, but reading line by line what's actually there.
You haven't got time for it, of course. There are already eighteen other things on your daily plate and at least this one's finished, you can see that. Except you can't see that at all. You see that it looks finished, and that's something else. Reading eight pages slowly takes twenty minutes and saps your team's attention. After that they still have to read the entire handbook and go and carry it to the clients. I hope they're not going to talk like an AI robot.
Four rounds cost us an extra two weeks. And the client gets an extra invoice.
Flowing AI language invites you to glide. Your eye shoots ahead, the sentences offer no resistance, and after three paragraphs you think you've read it. That doesn't happen with wooden text, which forces you to stop. So read your own text before you send it. Out loud, if you dare. You'll hear the repetition in paragraph four within ten seconds.
The same morning I'm thinking this, an email arrives that completely contradicts me. A client at HU in Utrecht has a group of people testing a new website. She delivers two Word documents. The first is written by hand, point by point, with screenshots included. Exemplary work. In the second document there's a technical analysis that a colleague of hers has had an AI produce. My client writes that she can't assess that piece herself.
I have every point from it checked in the code. No sitemap. No canonical. One page title for the entire site, in English, on the Dutch version as well. The questions in the FAQs aren't proper buttons, so anyone working with a keyboard can't get into them. It's all correct.
That report saves my team a day's searching and it finds something we need to know before going live. If I want to claim here that people are getting stupider from AI, I need to eat this example first.
Look at the difference between those two cases, because that's where my whole story lies. The tool is identical. The outcome is opposite. And the distinction doesn't lie in the machine.
The client writes that she can't assess it. That's precisely the sentence my client doesn't write.
One person sends through a proof and says it's a proof. The other sends through a proof and calls it his foreword.
So do you get stupider from AI? I think that remark oversimplifies things.
Michael Gerlich studies 666 people and publishes in early 2025 in the journal Societies that people who use AI heavily score lower on critical reasoning. The explanatory link is called cognitive offloading, outsourcing thinking work. The muscle you don't use gets weaker, and that applies to the muscle that judges as well.
The MIT Media Lab has participants write essays, with and without a language model, and measures brain activity whilst they're doing it. Anyone using the model is less active during writing, and performs worse afterwards when they have to do without. The study is called Your Brain on ChatGPT. Now the caveat. That's 54 participants. It's a preprint, and the researchers themselves warn against drawing overly large conclusions.
There's one finding in there that usually gets overlooked, and which changes everything. Education protects. People who are used to checking keep checking. Exposure to the technology predicts the damage less well than the user's habit.
So not: you get stupider from AI. Rather: AI rewards the habit you already had. Anyone used to checking gets an assistant. Anyone not used to it gets an excuse.
I work a lot for the fitness industry, a world built on appearance. There's nothing wrong with that, because that's the product. But writing has never really been a major craft there. So as soon as a machine arrives that produces full sentences effortlessly, things move quickly.
I know a gym in the west of the country whose website has been put together entirely by a language model. You see it immediately. Not in the text, because that's grammatically fine. You see it in the layout. The typeface is a size too small. There are coloured lines running across the page that lead nowhere. And some texts suddenly fall outside the nice rounded box. This is pure hallucination. There's not a single human logic in it.
It's a room furnished from a catalogue: thought has gone into each individual piece of furniture. But never once into the room. I see the same thing at an agency working at volume. The same AI fingerprint everywhere: the too-small typeface, the random colour accents, spacing that's calibrated to nothing. They're having a hard time, I know that, and I wish them the work. But what they're delivering is the proof.
Behind that lies something that concerns me more than the aesthetics. Anyone leaning entirely on a language model has little confidence in themselves. There's a difference between someone who says "help me do this faster" and someone who says "just do it, because I can't anyway". The first person uses a tool. The second person hides behind it.
My client with the bizarrely long foreword belongs to the second group, until I send him back. And then it turns out he can actually do it. That's the part that stays with me most. His own version is better. Not neater, not smoother, better. There's a signature in it.
The calculator gets this accusation too. Spell-check too. Every time the panic proves overdone, and I understand that someone might want to put me in that category now. Only something different is happening.
A calculator takes the operation out of your hands and leaves the judgement with you. You still decide which sum you're doing and whether the answer makes sense. A language model doesn't make that distinction. It delivers the operation and the judgement in one go, neatly packaged, and judgement is precisely the bit my profession consists of. Choosing what comes out. Seeing that paragraph four says the same thing as paragraph one. Knowing that a typeface needs to be two points larger. Feeling that a sentence doesn't sound like the client themselves.
That's not knowledge, that's taste, and taste only grows by using it.
So no, people aren't getting stupider from AI. They're getting lazy in the most expensive place.
Everything I don't enjoy goes to AI. That's not a confession. It's simply on one of my websites. Two of my team members are a language model. One coordinates and writes along, the other builds and repairs. They have a portrait, a profile page and a label that says what they are. With a wink. Because if you present this sort of thing deadly seriously, nobody believes you.
So if you're reading this piece as "old craftsman grumbles about new machines", you've got the wrong craftsman. The difference doesn't lie in whether I use it. The difference lies in what I do with it before it goes out the door.
Of everything that emerges from a language model here on any given day for clients, most of it ends up in the bin. Not because it's poor, because it rarely is. It goes because it doesn't match what we want to say, because there's too much of it, or because the tone isn't this client's. Throwing it away is the work. The machine writes, I decide. Anyone who reverses that gets exactly what my client gets: eight pages no one can bear. And something no one wants to read either.
This connects to an older habit, and it wasn't invented by the technology. People love to hide. Behind a report, behind a job title, behind a steering group, behind a forty-slide presentation in which nowhere does it say what anyone actually thinks. That's an existing instinct with a long service record.
What's new is that the hiding place is nearly free and delivered in four seconds. It used to take you two weeks to hide behind a report, and during those two weeks you'd bump into yourself a few times along the way. Now it costs you a prompt. So this isn't a new human failing. It's an old failing that can finally scale.
And the remedy is equally old-fashioned. Fill two A3 sheets. By hand, or typed if you must, but yourself. Not because handwriting is sacred, but because on sheet two you discover your argument doesn't hold. You never discover that when someone else writes it down, and a language model is someone else. Two A3 sheets produce less text than eight pages. They do produce something those eight pages lack: an opinion. And it's actually quite enjoyable to do.
The pendulum is swinging back. I'm already beginning to see it. Entrepreneurs who six months ago put together their own site with a chatbot are discovering there's nothing to rebuild. There's no system underneath. No grid, no typographic scale, no reason why something sits where it sits. Just individual decisions that happened to land beside each other.
They're coming back. Not because they've become more stupid from AI, and not because the technology disappoints, because it's excellent. They're coming back because they notice the technology has no opinion. And having a brand requires precisely that: an opinion.
My client takes four rounds to make something he could do in one. That's the price. Not that the machine writes poorly, because it doesn't. The price is that someone is still needed who dares to say it isn't finished, and that person increasingly has to come from outside.
I once asked on LinkedIn whether AI was making you more stupid yet. That question is already out of date. You used to check the proof yourself. That was called craftsmanship. Now it's called a second quote.
Michael Gerlich, AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking, Societies 15(1), 6, January 2025.
Nataliya Kosmyna et al., Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, MIT Media Lab, 2025.
Written by
Peter van der SteegePeter van der Steege is a brand strategist, designer and AI director. He builds brands for entrepreneurs and writes about what makes brands strong, from strategy to the role of AI and humanity. He lives and works in Groningen.
