Can AI write in your voice?

You have already tried it. You pasted in three things you wrote, asked for more in the same style, and read back something that was close. The vocabulary was right. The sentence lengths were about right. It used a phrase you actually use.

And it was not you, and you could not say which part was wrong.

The question gets answered badly in both directions. One camp says voice is a set of patterns and a model is a pattern machine, so obviously it can. The other says human expression holds something no system will ever reach. The first is an argument about capability. The second is an argument about souls. Neither answers the question you actually have, which is whether the thing you just pasted into your services page is any good.

There is research on this, and it lands somewhere neither camp expects.

In this article

• What happens to writing when a model helps
• Why the improvement and the sameness are one event
• Training it on your writing does not solve it
• What it is actually good at
• Common questions

What happens to writing when a model helps

In 2024 Anil Doshi and Oliver Hauser published an experiment in Science Advances. Around three hundred people wrote short stories. Some worked alone. Some received one story idea generated by GPT-4. Some received five.

Stories written with access to generated ideas were rated more creative, better written and more enjoyable than stories written without. The effect was not spread evenly. It concentrated among the writers who scored lowest on creativity to begin with. The most creative writers gained very little.

On its own that reads as a straightforward argument for using the thing.

Then the second finding. The researchers measured how similar the stories were to one another. The AI-assisted stories resembled each other more closely than the human-only stories did. Access to generated ideas raised the quality of any individual piece and narrowed the range across all of them in the same movement.

The authors call this a social dilemma. Each writer ends up better off. The collection ends up poorer. Applied to a market rather than a writing experiment, if everyone in a category reaches for the same assistance, everyone in that category improves and everyone in that category converges.

Which is the finding that matters for a website, because you are not writing in isolation. You are writing into a category where your competitors hold the same tools and feel the same pull towards them.

Why the improvement and the sameness are one event

It is tempting to read those as two separate results, one encouraging and one worrying. They are one mechanism observed twice.

A language model produces what is most probable given everything it has read, and probable means central. For a writer sitting below the centre of their category, being pulled towards that centre registers as improvement, which is what the study measured. For a writer sitting somewhere distinctive, being pulled towards the centre is a loss, which the study also measured, as the absence of gains among the most creative participants.

That explains something you have likely noticed without being able to name it. AI copy is rarely wrong. It is rarely badly written. It reads competent and faintly padded, because it includes everything a business like yours would say rather than the specific things you would say.

Voice runs on refusal. It shows up in what you decline to claim, the caveat you always add, the opinion you keep repeating that your competitors would never put on a website, and the point at which you stop talking. A model has no basis for refusing anything, because refusal requires a position and a position is the one thing it does not hold. Offered the chance to add a reassuring sentence, it adds it.

Which makes this the borrowed voice problem arriving faster and looking better dressed. Studying your competitors and absorbing their language used to take months. A model does it in four seconds, because it has already read them.

Training it on your writing does not solve it

The obvious response is that you have not given it enough of yourself. Feed it a proper corpus. Twenty posts, a style guide, a list of words you never use. Then it will be you.

There is a study on exactly this. Fiona Draxler and colleagues published it in ACM Transactions on Computer-Human Interaction in 2024, across two experiments with thirty and ninety-six participants. They named what they found the AI Ghostwriter Effect.

People given AI-generated text did not consider themselves its owners or authors. They also declined to publicly declare that AI had written it. Both of those held at once, which is an uncomfortable enough result on its own.

The finding that answers our question comes next. Personalising the generated text did not affect the ghostwriter effect. Tuning the output towards the individual did not make that individual feel like its author.

One thing did move ownership. Higher levels of the participant’s own influence on the text. Not more material fed into the system. More alteration of what came out of it.

So the corpus approach fails on its own terms. You can produce output that resembles your writing closely and still stand outside it, because ownership comes from exercising judgement on the text rather than from supplying raw material to something that will exercise judgement on your behalf.

The study has one more detail worth carrying. Participants granted more ownership to a supposed human ghostwriter than to an AI one, which produced a wider gap between felt ownership and declared authorship in the human case. Nobody is applying a consistent principle here. People are working out what they can live with, and the answer differs depending on who or what did the writing.

What it is actually good at

None of this makes the tool useless for writing. It makes it useful for the parts that are not voice, and the distinction is sharper than it first looks.

It works as a reader. Paste in something you wrote and ask what argument it is making. The summary tells you whether your argument survived contact with your own sentences, which is information you cannot get from rereading, because you already know what you meant.

It works on structure. Sequence is a solvable problem rather than an expression of who you are, so handing it over costs nothing. The order of your services page is not a claim about your character.

It works on the question of what is missing, which is a different question from what to say. Ask what a sceptical reader would object to and you get something genuinely worth having, because objections are patterns and patterns are what it holds.

It works as a draft you intend to dismantle, but only if dismantling actually happens. The Draxler finding puts a condition on this one. Ownership tracked the level of influence participants exerted on the text, so a draft you lightly tidy leaves you outside your own page. A draft you take apart and rebuild does not. The difference is not visible in the output and it is entirely visible in how you feel about the page six months later.

There is a practical version of the convergence problem too. If you are going to use it on anything public, the useful move is to give it your constraints rather than your samples. Samples ask it to imitate a surface. Constraints ask it to work inside a boundary you set. Telling a model what you refuse to say, which claims you will not make, and which words are banned gives it the one thing it cannot generate for itself, which is a position. The output still needs rebuilding. It starts from somewhere less crowded.

Which leaves the objection that none of this matters if no reader can tell. Detection is genuinely unreliable, and the honest position is that you will probably never be caught.

The Draxler result makes that objection less comfortable than it looks. Their participants declined to declare AI authorship while also not feeling like authors themselves, which is a description of writing under a claim you privately do not hold. Nobody catches you and you carry it anyway. That is the same structure as the trust gap that stops you automating client messages, where the cost lands on how you hold the relationship rather than on being found out.

There is also a straightforwardly commercial version. Convergence is measurable whether or not any individual reader identifies its cause. A visitor comparing four websites in your category does not need to detect anything. They need only fail to find a reason to choose you, and a page assembled from the centre of your category supplies no such reason.

What it cannot do is arrive at a position in the first place. Your positioning line, your About page, the sentence that says who you are wrong for. Those require somebody to decide, and the deciding is the whole content of the output rather than a preliminary to it. Delegating produces an answer without producing the decision, which is why the result feels borrowed even when no reader could tell.

That connects to the question of where AI stops paying in a business generally. The test there was what the output is being read as proof of. Voice is the sharpest case available, because a page written in your voice is a claim that a specific person wrote it, and no version of that claim survives being handed to something else.

If you cannot describe your own voice well enough to instruct anything, the obstacle is not prompting. Interrogating yourself harder will not clear it either, since self-report is the wrong instrument for that particular job. What reaches it is evidence. Go and read an email you sent a client you liked, written quickly with no thought of anyone else seeing it. Your voice is already in your sent folder, fully formed, waiting for somebody to notice what is in it. The Brand Mirror puts your answers next to your website and shows you the distance.

The honest answer to the question is that a model can produce something resembling your voice closely enough that most readers will never object. That is a different achievement from writing in your voice. The gap between the two is exactly the part you were hoping to skip.

Common questions

Can AI write in my brand voice if I give it enough examples?

It can produce something that resembles your writing. Whether that counts depends on what you want from it. Draxler and colleagues found that personalising AI-generated text did not make people feel like its author, while exercising more influence over the output did. Feeding it samples changes the surface. Rebuilding what comes back changes the ownership.

Why does AI writing sound generic even when it is well written?

Because a model produces what is most probable, and probable means central to everything it has read. Doshi and Hauser found that AI-assisted stories were rated better than unassisted ones while also resembling each other more closely. The quality goes up and the range narrows in the same movement, which reads as competent and slightly padded rather than as wrong.

Will using AI make my copy sound like my competitors?

If they are using it the same way, that is the direction of travel. The convergence effect operates across everyone using similar assistance, not within one business, so the risk is not that your page reads badly. It is that your page reads like the category.

What should I use AI for in my writing?

Structure, summarising your own draft back to you, and finding the objections you have not answered. Those are pattern problems and it holds patterns. Positioning, your About page and anything stating who you are wrong for require a decision, and the decision is the entire value of the output.