I use AI to help me write.
Quite a lot, actually.
Which is slightly awkward, considering that writing is part of my job, something I studied, and something I do in my spare time. Occasionally for an entire afternoon, without producing a single usable paragraph.
AI is considerably better at that last part. Or rather, at avoiding it.
It can produce a perfectly reasonable paragraph in seconds. Sometimes a better one than mine. Sometimes one I wish I'd written.
I don't particularly mind AI helping me find the right words. What bothers me is the possibility that, after enough suggestions, improvements and perfectly sensible rewrites, I might stop noticing when those words are no longer quite mine.
Not because AI has stolen my voice.
Because I've become very good at accepting its version of it.
So this is less a guide to keeping your writing voice when using AI than an attempt to work out where mine goes when I'm not paying attention. There will be practical advice. I just don't trust it to be the whole answer.
The short version: AI is very good at copying style, the part of your writing that’s easy to see. Your voice also lives in decisions: what you notice, what you leave out, what you’re prepared to stand behind. Use AI for friction and criticism rather than finished drafts, keep those decisions for yourself, and you can use it a lot without losing what makes the writing yours.
AI writing voice: style is the part that's easy to copy
Give an AI enough of my texts and it will learn that I like a short sentence after a long one. That I distrust neat conclusions. That I occasionally end a paragraph on something nobody asked for.
It will then produce texts that do all of that. Reliably.
Which is roughly the problem.
Style is what a text sounds like. Voice also includes what happens before there is any text at all. Which detail becomes the opening. What stays unexplained. When a contradiction is more interesting than its resolution. What you find funny. What you decide is worth writing about, even when a perfectly reasonable person might suggest doing something else with your Saturday.
AI is good at the first part. The second part is made of decisions. And decisions are exactly what becomes easy to hand over once the result already sounds fine.
Why AI rewrites sound polished and still lose your voice
Say I'm writing about the strange inability to stop working after work. I start with this:
I closed my laptop at six. At seven, I was still thinking about the email I hadn't answered.
I ask AI to make it more professional. It offers:
In today's fast-paced professional environment, maintaining a healthy work-life balance can be challenging, particularly as digital communication increasingly blurs the boundaries between personal and professional responsibilities.
There's nothing wrong with it. It is coherent and grammatical, and it could appear on several thousand corporate blogs without anyone noticing. It possibly already has.
It has also removed the laptop, the email and the quiet absurdity of having finished work while still sitting at your desk in your head. The original showed someone. The rewrite describes a category of people.
If I push back and work on it myself, I might end up here:
I closed my laptop at six. By seven, I had mentally answered the same email four times. None of the replies were particularly good.
Better? For some readers. More mine, certainly. It keeps the observation and lets the joke grow out of it, instead of swapping it for an explanation.
That's the trade I watch for. Not worse writing. Writing that becomes clearer, smoother and more professional in a way that is really just more conventional. One suggestion at a time, it's barely noticeable. After a few hundred, it's a style.
What research says about AI-assisted writing and ownership
At some point I wanted to know whether this was just my problem. Apparently not. But apparently it's also not quite the problem I thought it was.
In a study published by the ACM in 2025, Hwang and colleagues interviewed 19 professional writers and asked 30 readers to assess texts. The writers cared about authenticity, but not only in the finished text. They cared whether the process still expressed and developed who they were as writers. The readers, meanwhile, could not reliably tell which texts had been written with AI.
Which leaves an uncomfortable question. If readers can't see the difference, does it matter?
I think it does. Just not necessarily to the reader.
Two other studies point in a similar direction. Formosa and colleagues asked 602 participants to judge a scenario in which an author got help writing a short novel. The amount of help changed how much authorship, creatorship and responsibility people gave the author. Whether that help came from a human or from AI did not.
Pennanen and colleagues had 273 people write short stories with or without AI. Being able to edit the AI's text raised their sense of ownership. Iterative prompting helped too, but only when editing wasn't possible. Writing entirely by hand still produced much higher ownership than any of the AI conditions.
None of this tells you how much AI is too much. It does suggest that counting words is the wrong measure. What seems to matter is who was present when the decisions were made.
How I use AI without losing my voice
I don't have a rule that AI must never write a sentence. My professional work involves deadlines, briefs and search intent, and pretending otherwise would be a fairly expensive form of purity. What I do have is a rough division: AI helps me think before I write, and criticise after I've written. The part in the middle I try to keep for myself.
Most of my texts start with something incomplete. A remark in a meeting. A contradiction I can't quite explain. Occasionally a cow with an increasingly complicated relationship to modern employment. At that stage I ask AI where the tension is, not for a draft. Then I write the first version myself, because I often don't know what I think until I've written it. If AI hands me a polished draft too early, I end up editing an argument I never actually had.
Afterwards, I ask for criticism before I ask for a rewrite. When AI names a problem, I still have to solve it. When it rewrites everything, I can accept the solution without understanding it. Sometimes I put the awkward sentence back. Sometimes I delete it. Either way, I know why.
What this looks like in practice, from research to feedback, is in ChatGPT for Writers.
Where it matters most: AI and creative writing
In fiction, the flattening is easier to see. Take an opening like this:
He had been home for three days. His suitcase was still standing in the hallway, although there was nothing left inside it.
An AI rewrite will happily offer:
Three days after returning home, he was still struggling to settle back into his familiar surroundings, overwhelmed by a lingering sense of displacement.
It explains everything. That's the trouble.
The empty suitcase lets the reader work it out. Maybe he doesn't want to admit he's home. Maybe he has unpacked but hasn't arrived. Maybe the suitcase simply lives in the hallway now.
The rewrite treats that uncertainty as a defect. Often it's the point.
A useful AI editor would ask whether the suitcase comes back later, or whether the scene needs more context. What I don't want is for it to resolve the ambiguity because, from the outside, ambiguity looks exactly like a mistake.
In my short story Gott würfelt nicht (in German), three people's evenings run side by side, and nobody explains how they connect. A helpful assistant would have explained it. The story is largely about not doing that.
Not every unanswered question in a story is a weakness. Sometimes it is the story.
The Human Decision Map: where are you still making the decisions?
The more I use AI, the less useful I find the question "Did you use AI?" Almost everyone did, for something.
The more useful question is where you are still making the decisions.
That's what the Human Decision Map is for. It isn't a test of whether a text is human enough, and it isn't a scientific or legal assessment of authorship. It separates two things that are easy to confuse: AI contributing to a piece of work, and AI determining what that work becomes.
The map looks at two dimensions:
| Dimension | What it covers | Where it shows up |
|---|---|---|
| AI Participation | Where AI helps produce, explore or transform material | Research, Ideas, Structure, Drafting, Alternatives, Editing |
| Decision Authority | The consequential choices that set the direction and meaning of the finished work | Direction, Selection, Meaning, Final judgment, Responsibility |
Participation can be high without authority moving at all. If AI gives me five arguments and I reject all five, AI participated. It didn't direct anything.
The other extreme: AI picks the topic, the argument, the structure, the draft and the conclusion, and I make a few edits before publishing. That isn't AI-assisted any more. That's AI-led.
Here is the laptop example from above, on the authority side:
| Decision | In the laptop example | Who decided |
|---|---|---|
| Direction | Write about work that doesn't stop when work stops | Me |
| Selection | The corporate rewrite out, my own revision in | Me |
| Meaning | The absurdity of it, not "work-life balance" | Me |
| Final judgment | Which version goes into the text | Me |
| Responsibility | The sentence that gets published | Me |
On the participation side, AI was clearly involved. It offered an alternative and an edit. On the authority side, nothing moved.
Now imagine I'd accepted the rewrite. The word count would barely change. The map would.
That's the distinction I care about. Not how much AI is used, but who can still say no.
Try the Human Decision Map: ten decisions, about two minutes, no sign-up.
The part of writing I don't want to give away
I don't think the future of writing depends on proving that humans produce better sentences than machines.
Sometimes we do. Sometimes we don't. And sometimes the machine writes a perfectly good sentence that I simply don't want in my text.
Because the sentence was never the whole point. Writing is also how I find out what I think, what I notice and what I'm prepared to stand behind. AI can make that faster. It can make it better informed, and more honest about its weak spots. I welcome most of that.
But I'd like to be there when the interesting decisions are made.
Not because every sentence I write is precious. Some of them are terrible.
Mostly because I don't want to become so efficient at producing text that I forget why I wanted to write it.
Frequently asked questions
Can AI copy a writer's voice?
It can copy recognisable patterns: vocabulary, tone, sentence length, favourite moves. Whether it captures what the writer notices, and why they chose to say it, is a different matter. Style can be learned from examples. Voice also lives in decisions the examples don't show.
How can I use ChatGPT without losing my writing style?
Start from your own idea or draft. Ask ChatGPT for criticism and alternatives rather than full rewrites, and accept changes one at a time. Keep specific observations, rhythm and deliberate ambiguity, even when the polished version sounds smoother.
Does using AI make writing less authentic?
Not automatically. Research suggests that authenticity and a sense of ownership depend more on how involved the writer stays than on whether AI was used at all.
Should writers let AI write their first drafts?
It depends on the text. For routine professional content, an AI draft can be a reasonable start. For essays and fiction, the first draft is often where you find out what you think, which is a strange thing to delegate.
What is the difference between AI-assisted and AI-led writing?
In AI-assisted writing, AI contributes: research, ideas, drafts, alternatives, edits. The writer still decides direction, selection, meaning and the final text. In AI-led writing, AI makes those decisions and the human mostly approves. The difference is not how much AI is used, but who can say no.
About the author
Carsten Jan Weichelt is a Barcelona-based writer working across content strategy, creative writing and AI-assisted storytelling. He studied creative writing in Berlin and writes about technology, human judgement and authorship, including his Human-Led AI approach. His work includes the satirical storyworld Alma's World and the novel project Ich liebe das Leben. More about Carsten.