Writing With AI Is Making People Worse at Thinking. Here's the Fix.

Joan Didion said it plainly: "I write entirely to find out what I'm thinking." Not to report a finished thought. To find one. The sentence on the page isn't a transcript of the idea, it is the idea, worked out in real time by the friction of having to make it grammatical and specific and true.

If that's right, and most working writers feel in their bones that it is, then something uncomfortable follows. When you hand the writing to a model, you're not just outsourcing the typing. You're outsourcing the finding-out. And writing with AI, done the lazy way, is quietly making a lot of people worse at the one thing writing was secretly for: thinking.

Writing is thinking, not the packaging of thinking

There's a popular model of writing that treats it as a translation job. You have a thought, fully formed, somewhere in your head, and writing is the chore of converting it into words. On that model, AI is a pure win. Same thought, less typing.

But that model is wrong, and any honest writer knows it. You almost never have the full thought before you write it. You have a fog, a direction, a feeling that something is there. The thinking happens in the writing, sentence by sentence, as you're forced to commit. You discover you don't actually believe your own opening. You realize the example doesn't fit. You find the better point hiding three paragraphs down. That's cognition, and it only happens because the writing made it happen.

Hand the sentences to a model and the fog never has to resolve. You get clean prose built on a thought you never actually finished having. It reads fine. It's hollow, and often you can't tell, because the polish hides the absence.

The cognitive-offloading problem, in plain terms

Cognitive offloading is the well-studied habit of pushing mental work onto external aids. Calculators, GPS, the phone number you no longer remember because your phone does. Mostly this is fine, even great. Nobody needs to memorize phone numbers.

The catch is that offloading has a cost when the thing you're offloading is the thinking you actually wanted to keep. There's a tidy line of research on GPS: heavy reliance is associated with worse spatial memory and weaker navigation skill over time. You stop building the mental map because the app holds it for you, and then one day the app is wrong and you have no map at all. The relationship between using the aid and losing the underlying ability is direct.

Writing-as-thinking is the same shape, with higher stakes, because the ability you're offloading isn't finding a restaurant. It's reasoning. When you let AI generate your argument, your scene, your explanation, you skip the part where your brain wrestles the idea into a shape it can defend. Do that often enough and the wrestling gets harder, because you've stopped doing it. The muscle that turns fog into a real thought atrophies, and it's not a niche muscle. It's most of what we mean by thinking clearly.

You can feel this happening if you watch for it. The tell is the moment you reach for the model not because you're stuck on a word, but because you're stuck on the idea and you'd rather not sit in the discomfort. That discomfort is the work. Skipping it is the offload.

Why the polished output hides the rot

The genuinely insidious part is that bad thinking used to look bad. Muddled reasoning produced muddled prose, and the bad prose was a warning sign, to your reader and to you. Clunky writing was diagnostic. It told you the thought wasn't done.

AI severs that signal. Now half-baked thinking comes out fluent. The model smooths the surface so you can no longer hear that the foundation is missing. A confused idea reads like a confident one. You lose the early-warning system precisely when you need it most, and so do your readers, who can no longer use clumsiness as a clue that someone didn't think it through.

This is why "but the output is good" misses the point entirely. Output quality was never the only thing the writing was producing. It was also producing you, the person who can think the thought. That product is the one quietly disappearing.

The fix: make AI interrogate your thinking, not replace it

Here's the part the doom takes always skip. The same tool that can offload your thinking can also sharpen it, and the difference is entirely in how you use it. You can point AI at your reasoning instead of at your typing. Used that way it's not a substitute for thinking. It's a pressure test for it, which is the entire premise behind treating the model as an AI writing coach rather than a generator.

A few moves that keep the thinking on your side:

• Draft your real thought first, badly, in your own words. Get the fog onto the page yourself. The bad draft is where the thinking happens, so this step is non-negotiable. Only then bring the model in.
• Ask it to attack, not to write. "What's the weakest link in this argument?" "What would a smart skeptic say?" "What am I assuming here that I haven't earned?" These don't replace your reasoning. They stress it, and you get stronger answering them.
• Make it surface what you're avoiding. "What's the obvious objection I'm dodging?" The model is genuinely good at naming the thing you flinched away from, and naming it forces you to actually deal with it.
• Never accept a conclusion you didn't reach yourself. If the model hands you a point that sounds smart, the work is to figure out whether you believe it and why. The reasoning has to pass through your head, not around it.

Notice that every one of these keeps you doing the cognitive labor and uses the model to make that labor harder, not easier. That's the whole inversion. We've made a parallel argument about outsourcing your taste to a language model, and it's the same disease in a different organ: the moment you let the model's default stand in for your own judgment, the judgment stops developing.

Build the friction back in on purpose

The deeper fix is structural, not just prompt-by-prompt. Set your process up so the model can't sit in the thinking seat by default. Keep a no-AI first pass on anything that actually requires you to figure out what you believe. Use the tool downstream, as a sparring partner, after the thinking is at least started. This is what what Polyz does is built around: keeping you generating and reasoning, and putting the AI in the chair that pushes back.

The goal isn't to write slower for its own sake. It's to protect the only part of writing that was ever really for you. The book, the essay, the argument: those are byproducts. The thinking was the point. Keep that, and AI is the best sparring partner you've ever had. Lose it, and all the fluent output in the world won't give it back.

---

Polyz keeps you thinking on the page and uses AI to interrogate your work, not write it for you. Start a free trial.