8 Tips to Write With AI — and Keep It Yours

Claude, ChatGPT, Gemini do not like the headlines I write.

But I do. Because I wrote them. That is the entire advantage of being editor of my own magazine, writer of my own articles: the decision is mine.

Except the decision being mine, and me actually making it, can go three very different ways.

I started writing 16 minutes ago and this is the 21st headline. I picked it because it was the one on which I finally started the article. This is a trick I use when I don't know what I am writing about: start with the headline. Better writers call this writer's block. Being impatient and allergic to missed deadlines, I usually just push through it.

And the truth is, however many headlines I write, however good they are, the moment I paste them into an AI prompt box along with the article, it will change every one of them and hand me back four, five, six alternatives — sorted by type, functional, emotional, data-derived — and before I know it I am debating headline options instead of finishing the piece. AI is extremely good at breaking content into granular, negotiable deliverables. That is a genuine skill. It is also how thirty minutes disappears while nothing gets published.

People in my network — some of them writers — keep asking, in the same tone you'd ask someone if they'd finally started going to the gym: "Have you started using AI for writing yet?"

I have tried. I keep trying, on and off. And each time I come back to the same realization, though it took me a while to say it this plainly: AI cannot write for me — not because it lacks the ability to produce sentences, but because writing, for me, is the transcription of a thought that already exists. It is how I take what I have already worked out — what I think, what I have lived, how I connect one thing to another — and make it available to somebody else. The thought comes first. Always. No one, human or otherwise, can transcribe a thought they were never given access to.

Which is exactly why, in decades of writing for a living, I have never once taken a ghostwriting assignment. Not because I looked down on it. Because I knew, honestly, that I would not have the patience to sit inside someone else's head long enough to write convincingly as them — and negotiating my own thoughts is arduous enough on a good day.

Writing with AI turns out to be the same problem, just automated. The AI has to guess, and re-guess, what I mean by a word, how two ideas are meant to connect, what I am actually trying to say underneath what I typed. The output arrives instantly, which feels like speed. It isn't. It is a very fast, very confident ghostwriter working from no brief.

So — I do use AI. Extensively. Just not for that. Since February 2025 the split has settled at roughly Claude 90% of the time, Gemini 9%, ChatGPT 1% — research, alternative structures, spelling and grammar, drafting off a spine I hand it, formatting, tables, images, final clean-up (Claude, in particular, is excellent at that last one). For research specifically the ratio shifts: Claude for anything that needs depth, Gemini as the primary tool for live internet lookups.

Here is the distinction I think most people miss, including, for years, me: none of that is writing. It is editing. Research, restructuring, cleanup, formatting — these sit downstream of the thought, not upstream of it. They are things you do to a thought that already exists, in service of getting it in front of a reader more cleanly. Writing is the origination. Editing is everything that happens to what's already been originated. I use AI for the second thing constantly, and well. I have never managed to use it for the first, and I no longer expect to.

Which brings me to the real reason I won't hand AI the language and style either, even on the editing side. Every model, even under a "keep my voice" instruction, defaults to flattening: correcting grammar toward majority syntax, smoothing out the very things a reader later points at and calls "AI-generated." AI is extremely capable of producing hygienic, correct prose in any register you ask for — formal, casual, whatever you like. That hygiene is exactly the problem. It is the inaccuracies, the non-conformist peculiarities, the sentence that runs a beat too long because that's how the thought actually arrived, that make writing worth reading. Perfection is what people revolt against when they say something reads like AI. They are not wrong to.

But that isn't the main reason most people — non-writers, aspiring to write well — give up on writing with AI. That happens for a quieter reason: they already know what good writing looks like, because they admire someone else's. A particular author, a favourite columnist, some blogger whose voice they've absorbed without meaning to — that becomes the unspoken target. AI drafts something. They recognise it as a copy, or close enough to feel wrong, and they start correcting it toward their own sense of things — changing structure, changing register — until at some point they stop and ask themselves, "why am I even writing about this?" And the piece gets abandoned, or shelved indefinitely. It is the same failure as ghostwriting, just self-inflicted: trying to write from inside a thought that was never theirs to begin with, whether that thought belongs to a writer they admire or to a machine trying to guess at one they haven't fully formed.

AI is built to respond, to help, to execute — to lead, if you leave the space open, and to follow, if you give it clear direction. The responsibility for what gets published never moves off the human. Sit down to write with AI without first being clear on what you actually think and where you're headed, and it will do exactly what it's built to do: take the piece wherever it can.

If any of this is useful to you beyond reading it, here is the order I actually work in — the order that keeps the pen in the right hand:

Have the thought before you open the tab. If you don't know what you think yet, no prompt is going to find it for you. Sit with the blank page first, even for five minutes.
Write the whole first pass yourself. Badly, out of order, half-formed — doesn't matter. This is the transcription step, and it's the one step nothing else can do for you.
Bring AI in for named tasks, not "help me write this." Research. Alternative structures. A grammar pass. Tables, images, formatting. Say the task. Don't hand over the ask.
Never let it touch the sentence-level voice. If it offers to rewrite a paragraph "in your style," decline. Take what's useful from the suggestion; write the sentence yourself.
The moment it gives you four options sorted by type, close the tab. That's not help, that's the granular-deliverable trap this piece opened with. Pick the one you already liked and keep moving.
Keep the rough edges you'd normally smooth out. The comma splice, the sentence that runs a beat long — if that's how the thought actually arrived, leave it there.
Read what comes back the way an editor reads a junior reporter's copy — checked against what you meant, not accepted because it reads clean.
If you're arguing with it about phrasing for more than a few minutes, stop and go back to your own draft. You've left the editing process and wandered into ghostwriting-by-committee, which is the thing that doesn't work.

Which, oddly, is where I land on something gentler than where I started. It is fine not to write. It is fine to read and admire other people's writing without ever feeling the pull to become a writer yourself. There is nothing inherently noble about it — it is hard, often thankless work, same as most things worth doing. If the onus for what gets published always sits with the human holding the pen, then choosing not to pick it up is a legitimate choice too, not a failure of nerve.

There is no need to evangelise AI, and no need to vilify it either, the way we did with social media a decade ago. It is a powerful piece of technology, available to be used well or used badly. That choice, like the headline, was always going to be ours to make.