There’s a scenario most people who use AI for writing have hit at least once and quietly ignored: you paste in the AI draft, start editing, and thirty minutes later realize you’ve rewritten every single sentence. The draft didn’t help. It actively slowed you down because you spent your energy reacting to someone else’s bad choices instead of thinking clearly about what you actually wanted to say.
That’s worth sitting with, because the default advice — “just let AI get you past the blank page” — treats all blank pages as the same problem. They’re not.
When the blank page is actually the right tool
There are two different things that make writing feel hard. One is not knowing what you want to say. The other is knowing exactly what you want to say but finding it effortful to get into words. AI is genuinely useful for the second one. It’s often counterproductive for the first.
If you’re writing something where the thinking is the work — an explanation of a complex idea, an argument you’re still working out, a piece where your specific point of view is the whole point — generating a draft first tends to collapse your thinking prematurely. You get a version of the topic that sounds plausible, and then you’re stuck either accepting its framing or fighting it. Neither of those is the same as actually developing your own take.
This shows up in a particular way with opinion pieces or anything that’s supposed to sound like you. The model will produce a structurally reasonable draft with a thesis, some supporting points, a tidy close. But it’ll be the thesis the model thinks is most defensible on the topic, not the one you’d actually argue. Editing that into your real position takes longer than writing from scratch would have, because you’re constantly overriding choices you didn’t make.
The cases where AI drafts genuinely earn their keep
This isn’t an argument against using AI to write. It’s an argument for being more deliberate about when you reach for it.
AI drafts work well when the format is heavily constrained and the content is mostly factual. Product descriptions, FAQ answers, templated emails, standard explainer structures — these are cases where the model’s tendency to produce “the most statistically normal version of this thing” is actually what you want. You’re not trying to express a unique perspective; you’re trying to produce a correct, readable instance of a known format.
They also work well when you have something specific to give the model: an outline you’ve already thought through, bullet points of the actual claims you want to make, a rough draft that needs restructuring rather than invention. When your thinking is already done and you just need words around it, a model is fast and competent. The problem is specifically when you’re asking the model to do the thinking for you and then trying to retroactively own the output.
A more useful alternative: talk before you draft
One pattern that actually helps: before asking for a draft, use the model as a thinking tool. Describe what you’re trying to write and ask it to push back on your framing, ask clarifying questions, or play devil’s advocate against your main point. This gets the model’s pattern-matching working in your favor — it’ll often surface an objection or an angle you hadn’t considered — without surrendering authorship of your actual argument.
Then write the draft yourself, or write an outline first and ask the model to fill in sections you already know the shape of.
It’s slower than just saying “write me a post about X.” But the output doesn’t need surgery afterward.
The editing trap
There’s a subtler version of this problem worth naming. Even when you’re editing an AI draft you do want to keep, the act of editing it tends to make you conservative. You fix the things that are clearly wrong, but you accept the structure, the order of ideas, the implicit assumptions baked into the draft’s framing. You end up with something polished that still isn’t quite what you would have written, and you often can’t articulate exactly why it feels slightly off.
This is a known problem in writing more broadly — editing someone else’s draft is easier than writing from scratch but produces a different kind of output than writing from your own outline would. With AI drafts it’s the same issue, just with the additional quirk that the “someone else” has very confident, fluent opinions about what a piece on this topic should say.
Worth asking: how much of your editing time is spent fixing things the model chose, rather than refining things you chose? If the answer is most of it, the draft isn’t actually helping.
A practical heuristic
Before generating a draft, try writing one sentence: the actual claim you want to make, in your own words, as directly as possible. Not a description of the topic — the argument. If you can write that sentence, you probably have enough to write the piece yourself with AI helping you expand specific sections. If you can’t write that sentence yet, generating a full draft won’t solve that problem. It’ll paper over it with someone else’s argument.
The blank page is uncomfortable. But sometimes it’s doing useful work — it’s the pressure that forces you to figure out what you actually think. Filling it with a plausible-sounding draft before you’ve done that figuring is one of the more effective ways to avoid thinking while feeling productive.
Try writing that one-sentence claim before you hit generate. If it takes more than a minute, that’s the actual work — and no draft is going to do it for you.
