Keeping Your Voice When the Machine Drafts

Anyone who has read enough generated prose can recognise it, and the reasons are mechanical rather than mysterious.

A model produces, at each point, something like the most probable continuation given everything before it. Aggregated over a whole piece, that pulls hard toward the centre of everything it has read. The result is prose with identifiable properties.

Balance. Every claim arrives with its counterweight. It is not just X, it is also Y. On the one hand, on the other. This reads as fairness and is frequently evasion, because the writer has not decided anything.

Summary sentences that restate what was just said. Openings that announce what the piece will do rather than doing it. Closings that reach for significance.

Tricolons, because three-part lists are pleasing and common. Elevated connectives. A certain smoothness where nothing is awkward and nothing is surprising.

And an aversion to the flat declarative. Generated prose hedges, because hedged statements are more often true across a large corpus.

None of these are errors. Each is a reasonable local choice. The problem is that they are the same reasonable local choices every time, which is what makes the aggregate recognisable.

And this is the mechanism behind the marketing observation from elsewhere in this catalogue: a model produces the typical version of the thing requested. In writing, typical is the one quality that has no value, because a reader chooses a writer for what only that writer would have said.

Voice gets discussed as though it were a decoration applied to finished prose. It is not, and knowing what it consists of tells you which parts a tool can touch.

What you notice. A writer's attention is idiosyncratic. Two reporters at the same event produce different pieces because they found different things interesting, and that difference is upstream of any sentence.

What you are willing to say. Voice largely consists of commitment. A writer with a voice makes claims, declines to balance every statement, and accepts being wrong in public. Hedging everything produces prose with no voice by construction.

Your reference frame. The things you compare to, the examples you reach for, the fields you borrow from. These come from a specific life and reading history.

Rhythm and diction. The sentence lengths you favour, the words you would never use, where you place the weight. This is the most visible layer and the shallowest.

And what you leave out. Much of style is omission, and omission requires knowing what the reader can supply.

The first three are judgement, and they precede writing. The fourth is craft, and it is learnable. The fifth is judgement again.

So a model can imitate the fourth layer given examples, and it can do so well enough to be useful. What it cannot supply is the first three, because they come from having a particular mind pointed at the world. When a writer outsources the whole draft, the layers that carried the voice were the ones never delegated to prose in the first place.

The same tool used in two directions, with opposite effects on voice.

The ghostwriter direction: you supply a topic and instructions, the model supplies prose, you edit it. What arrives is centre-of-corpus writing, and editing pulls it only partway back, because you are correcting toward your voice from a starting point that is not yours. Repeat this for months and your sense of your own register drifts, because the thing you read most in your own byline was not written by you.

The editor direction: you write the draft, badly and fast if necessary, and the model works on what you produced. Tighten this. Where is this unclear? What am I assuming the reader knows? Which paragraph is doing no work? Argue against my third point.

In the second direction every word originated with you, so the voice is yours by construction, and the model is doing what a good line editor does: finding the flab, the ambiguity, the unearned assertion.

The asymmetry worth noticing. Both directions produce a comparable finished piece today. Only one of them leaves you a better writer next year, because in the editor direction you are still doing the composing, which is how composing improves.

So the question is not whether to use the tool. It is which way the text is flowing.

The editor direction only works if you ask for something more useful than make this better. Specific requests that reliably improve a piece.

Where is this unclear to a reader who does not already know this subject? Writers systematically overestimate shared context, and this catches it.

Which paragraph could be cut without losing anything? Usually there is one, and usually the writer suspected it.

What am I assuming that I have not established? This finds the missing step in an argument, which is the most common structural fault.

Argue against my central claim as strongly as you can. Useful because the strongest counterargument is what a hostile reader will bring, and it is better to meet it in the draft.

Where does this get slow? A model is a reasonable proxy for attention, and a passage flagged as slow usually is.

What is this piece actually about? Ask this before you have said, and compare the answer to your intention. A mismatch means the piece is not doing what you think.

What did I put in the fourth section that should be in the first? Buried leads are endemic and hard to see in your own work.

What none of these do is generate prose. Each returns a diagnosis you then act on yourself, which keeps the composing where it belongs, and each is a question a good editor would ask.

The general shape: ask for problems, not solutions. A problem you fix yourself teaches you something; a solution handed to you does not.

The most common defence of the ghostwriter direction is that a generated first draft solves the blank page. It is worth taking seriously and then examining, because it contains a real observation and a mistaken conclusion.

The real observation. Starting is genuinely hard, and having something on the page reduces the resistance. That is true and it is why writers use all sorts of tricks to get moving.

The mistaken conclusion is that any text serves that function equally.

What makes a blank page hard is usually not the absence of words. It is that you have not yet decided what you think, or what the piece is, or where it starts. Those are thinking problems, and text that papers over them does not solve them. It hides them under fluent prose, which is worse, because now the piece appears to exist and the unresolved decision is buried inside it.

And there is a second cost. A draft in front of you anchors. Once a structure exists on the page, most writers edit within it rather than discarding it, even when the right move was to start somewhere else entirely. That anchoring is well documented in judgement research generally and it applies straightforwardly here.

What actually addresses the blank page without those costs. Write badly on purpose and fast, knowing it is disposable. Talk it through, out loud or into a transcription, since most people can explain a piece they cannot yet write. Or use the model to interrogate rather than to produce: what are the three ways this piece could be structured, what is the strongest version of the argument.

Each of those gets you moving without handing over the decision that was actually blocking you.

Everything so far applies to journalism. Commercial content writing faces the same dynamic with the economics pointing the wrong way, and it deserves its own treatment.

The pressure is structural. Content is frequently commissioned by volume and priced per piece, and search-driven publishing rewarded quantity for years. That combination makes generated output enormously tempting, because the client is buying words and the tool produces words.

What has changed is the value of those words. When producing competent, generic content on any topic costs almost nothing, competent generic content is worth almost nothing. The floor rose for everyone, and everything at the floor became commodity.

Search engines have also shifted toward rewarding demonstrable expertise and original information over comprehensive coverage of a topic, which is a direct response to the flood of adequate generated material.

So the strategic position for a content writer inverts. The defensible work is exactly what a model cannot produce: original research, actual interviews with named people, first-hand testing, proprietary data, opinions someone will disagree with, and specific experience.

A piece containing an interview with a practitioner, a test you ran, or a number nobody else has cannot be generated, because the input does not exist in any corpus.

Which points to the honest advice for anyone in this field. The route through is not producing more, faster. It is producing things that require having done something, and the writers who struggle will be those whose output was already synthesis of what was findable.

Whether to tell readers about AI use divides into cases that are frequently conflated, and the distinction is about deception rather than about tooling.

Where disclosure is owed. Where the reader would change their assessment of the piece if they knew. Content substantially generated rather than written. Any synthetic image, audio or video presented as documentary, which is both an ethical matter and increasingly a legal one, since the EU AI Act's Article 50 requires machine-readable marking of synthetic media with that obligation applying from 2 December 2026. A synthesised voice reading a piece as though it were the author. And an automated system a reader might take for a person.

Where it is not owed. Transcription of your own interviews. Line editing your own prose. Research assistance you then verified. Headline variants you chose between. Nobody expects a byline to certify that no software touched the piece, and no reader assumed a spellchecker was disclosed.

The test that separates them. Is the reader being misled about something that would change how they weigh the work? Generated prose under a human byline meets that test, because a byline is a claim of authorship. A tool that helped you tighten your own paragraph does not.

What matters more than any policy is that the standard is public and consistent. Many outlets have now published theirs, and having one is largely what protects an operation when a specific case gets scrutinised.

And note the direction of the obligation. It runs to readers, not to the profession. The question is always what a reader would want to know, not what colleagues would forgive.

A closing point that is speculative in its magnitude and not in its direction, and it is worth stating plainly.

Writing is not the transcription of finished thoughts. For most people it is how thinking gets finished. You discover that an argument does not hold when you try to write the sentence that connects two parts of it. You find out what you actually believe by attempting to state it and recoiling. The struggle in composition is diagnostic: it is usually telling you that you do not yet understand the thing.

Which means outsourcing composition does something other than saving time. It removes the step where the thinking was getting checked.

This is why the ghostwriter direction is more costly than it looks. The output is comparable. The process that produced it skipped the part where you would have noticed the argument was weak, and generated prose is unusually good at making a weak argument read smoothly, because smoothness is what it optimises.

And there is a compounding version. A writer who drafts by hand for ten years gets better at thinking in prose. A writer who has always had a draft handed to them has not practised the thing, and will be less able to tell whether the draft is any good, which is precisely the judgement the whole arrangement depends on.

None of that argues for refusing the tools. It argues for keeping the composing and delegating everything around it, which is what this lesson has been describing throughout. Use it as an editor, a transcriber, a researcher and a critic. Write the sentences yourself, because writing them is how you find out whether you have anything to say.