There’s a workflow that’s become almost universal among people who write regularly and use AI: hit a blank page, type a rough prompt into ChatGPT or Claude, use whatever comes out as a starting point. It’s faster than staring at a cursor. The output is coherent. It saves the worst part of writing, which is the first sentence.
I do this too. And I’ve noticed something uncomfortable about what happens when I do it too often.
The problem isn’t the quality of the draft. It’s that struggling with a blank page is where a lot of the actual thinking happens — and when you skip it, you skip the thinking too.
What the blank page is actually doing
Writing from scratch forces you to answer a question most people never consciously ask: what do I actually think about this? That sounds obvious, but it isn’t. Most of the time we have a vague sense of a position, and putting words on a page is what sharpens it into something specific. The friction is the mechanism. You find out what you believe by being forced to state it.
When you hand that friction off to a model, it generates a plausible, structured version of the topic — shaped by whatever pattern of writing on that subject dominated its training data. That output isn’t your thinking. It’s a weighted average of how this kind of thing gets said. You then edit it, which means you’re reacting to someone else’s structure instead of building your own.
This matters less for some tasks than others. Confirmation email? Generate away. Product description where all the facts are fixed? Fine. But for anything where your actual perspective is the point — an essay, a client proposal where your reasoning is what the client is paying for, a piece you need to stand behind — the shortcut has a real cost.
The compounding effect nobody talks about
Skills atrophy when you stop using them. Not controversial. What’s less discussed is how fast it happens with writing specifically.
Writing is one of those skills where the gap between “can produce something” and “can produce something distinctively mine” is almost entirely built through repetition of the hard version. Every time you work through a blank page, you’re reinforcing your own voice, your own default structural choices, your own sense of when an argument is tight versus when it’s hand-wavy. Skip that step often enough and the skill doesn’t sharpen — it drifts.
I noticed this running an automated blog pipeline that generates posts with no human writing them from scratch. The system works well for its purpose. But I’m under no illusion that operating it is teaching me anything about writing. It’s a production system. Treating it as a writing practice would be like saying you got better at driving because you took a lot of taxis.
Where the trade-off actually lands
The honest version of this isn’t “AI drafts are bad.” It’s more specific.
If your goal is throughput — more content, faster, for a context where volume matters and individual pieces don’t need to carry personal weight — then AI-first drafting is a legitimate production choice. No argument there.
If your goal is to get better at writing, or to produce work where your specific reasoning and voice are the actual value being delivered, then AI-first drafting is working against you. Not because the output is bad. Because the process of generating it isn’t building anything in you.
There’s a middle approach that works better than either extreme: write your own rough version first — even badly, even just bullet points of what you’re actually trying to say — and then use AI to react to it. Ask it to poke holes in your argument. Ask if there’s a counterpoint you’re not addressing. Use it as an editor and a critic rather than a ghostwriter. The structure stays yours. The thinking stays yours. The AI is doing the part it’s genuinely better at, which is pattern-matching against a huge surface area of related ideas to catch what you missed.
A quick practical note on prompting for this
If you want to use AI as a critic rather than a drafter, the prompt structure matters. Vague requests get vague pushback. “Here’s my argument — tell me the strongest objection someone who disagrees would raise, and whether I’ve actually addressed it” gets you something useful. “Give me feedback on this” gets you a list of compliments with one polite suggestion at the end.
The more specific you are about what kind of problem you want it to find, the more useful the response. Ask for the weakest link in your reasoning. Ask whether the conclusion actually follows from what you set up. Ask if there’s a simpler way to say the core claim. These are all things AI is genuinely good at — not because it understands your argument deeply, but because it’s seen enough arguments to recognize what structural failure usually looks like.
The caveat worth keeping in mind
None of this means AI-generated drafts are inherently lower quality than human-generated ones — for many tasks they’re clearly not. And the skill-atrophy concern doesn’t apply equally everywhere. Nobody worries that using a calculator is costing them their long-division ability, because long division isn’t worth preserving for its own sake.
Writing is different mainly if thinking-through-writing is how you actually figure out what you think. For those people, outsourcing the blank page is outsourcing the thinking. So: are you using AI to do a task more efficiently, or to skip the part of the task that was doing something for you? That’s the distinction worth being honest with yourself about.
