AI Copywriting That Keeps Your Voice

AI copywriting is the use of generative models to produce marketing copy: ads, landing pages, emails, product descriptions, sales pages. It works, it is fast, and its default output will quietly erase the thing that made your copy worth reading. That is not a flaw in the tools. It is what the tools are built to do, and understanding why is the difference between using them and being flattened by them.

A language model is trained to predict the most likely next word. Averaged across everything ever written about your subject, the most likely next word is the most ordinary one. Voice is the opposite: a consistent, deliberate deviation from the average. So a model does not fail to reproduce your voice by accident. It regresses towards the mean because regressing towards the mean is the job.

The industry has noticed. Among content teams with AI guidelines, 59% say those guidelines explicitly emphasise preserving brand voice, and 49% report that their leadership requires human review before anything is published (Content Marketing Institute). Meanwhile 77% of marketers using generative AI use it for creative development, the single most common application (Gartner, CMO Spend Survey 2025). A lot of copy is being produced by a system that averages, inside organisations that have noticed the averaging and have not yet worked out what to do about it.

What voice actually is

“Brand voice” is usually documented as a list of adjectives, which is why the documentation does not work. Nobody has ever fixed a flat paragraph by being told to sound “confident, approachable and human.” Voice is made of four concrete things, and all four can be written down precisely enough for a model to follow.

Diction. The words you reach for and the register you hold. Whether you write “buy” or “invest in”, “problem” or “challenge”, “we think” or “it is widely believed”.

Rhythm. Sentence length and how it varies. Whether you run long and subordinate, or short and declarative, or deliberately alternate. This is the most audible part of a voice and the least often documented.

Refusals. What you will not say. The claims you decline to make, the hype words you never use, the flattery you skip. A voice is defined at least as much by its exclusions as its inclusions, and exclusions are the easiest thing to specify and the most reliably ignored.

Stance. The position underneath the words. Whether you are a guide, a challenger, an insider, a sceptic. Stance is why two writers can use identical vocabulary and sound nothing alike.

A model can be steered on all four. It just cannot infer them from three adjectives.

Why “match my tone” does not work

The standard instruction, write this in our brand voice, friendly but professional, fails for a simple reason: those words describe a very large region of language space. Almost anything can be friendly but professional. The instruction does not narrow.

What narrows is examples. A model shown six paragraphs of your actual best writing will imitate the rhythm and diction far more accurately than one given a page of adjectives, because the examples contain the information the adjectives only gesture at. Negative examples narrow even harder: showing what you would never write removes a whole region of output that adjectives leave wide open.

This is the practical asymmetry worth knowing. Descriptions of voice are cheap to write and weak in effect. Examples are more work to assemble and do almost all of the steering.

A voice specification that actually holds

The teams keeping their voice intact have generally built the same artefact, whatever they call it. It is short, and it is made of evidence rather than aspiration.

• Six to ten paragraphs of your own best work, chosen because they sound most like you, not because they performed best.
• A banned list, written as specific constructions rather than vague warnings. Not “avoid jargon” but the actual phrases: “in today’s fast-paced world”, “unlock the power of”, “it’s not just X, it’s Y”, “game-changer”, “seamless”. Constructions, not topics.
• Rhythm rules stated as constraints a model can check itself against: maximum sentence length, how often a short sentence must appear, whether you permit rhetorical questions.
• Three before-and-after pairs: a flat version and your rewrite, which teaches the edit rather than the style.
• The stance in one sentence: who you are to the reader, and what you refuse to pretend.

That document is short enough to paste into any tool and specific enough to enforce. It is also, not incidentally, the thing that survives when the tool you are using changes, which is why it belongs to you rather than living inside a vendor’s brand-voice feature.

Where AI genuinely helps with copy

None of this is an argument for writing everything by hand. The model is excellent at several things that used to eat a copywriter’s day.

Variants and testing: producing fifteen headline options is a volume problem, and volume is what the machine is for. Compression: cutting a passage to half its length while keeping the argument is a task models do well and humans find tedious. Structure: turning a messy set of notes into an ordered outline. Repurposing: taking one finished piece into the formats it needs to exist in. And the blank page, which a model removes entirely.

What it should not own is the opening line, the central argument, and the close. Those three carry the persuasion, and they are exactly where the averaging does the most damage, because a generic opening is indistinguishable from every other generic opening the reader has already skipped past today.

The division is not “AI writes, human edits.” It is closer to the reverse: the human sets the position, the structure and the sentences that carry the weight, and the model does the volume around them. This is the same split that decides whether AI-assisted content differentiates or dissolves, and it applies with more force in copy, because copy is where a generic sentence costs money immediately rather than eventually.

The test for whether your voice survived

Read the finished copy and ask whether a competitor could publish it under their own logo with no edits. If they could, the copy describes a category rather than a company, and you have paid to produce something that works equally well for the people you are competing with.

The second test is quieter and more useful: read it aloud. Voice is an audible property, and the ear catches the averaging faster than the eye does. The sentence that sounds like it came from nowhere in particular usually did.

Keeping a voice is not a matter of writing everything yourself. It is a matter of deciding, in advance and in writing, what yours consists of, so that a system built to produce the average has something specific to aim at instead. That specification is a small asset, and like the rest of the context layer that makes AI output yours, it compounds quietly and belongs to you.

If you want to build your own voice specification and the workflow around it, The Worldshaper Workshop is Daniel Förster’s daily build sessions, run by Thane Alaric: watch the work happen, then build it yourself. See The Worldshaper Workshop.

Frequently Asked Questions

What is AI copywriting?

AI copywriting is using generative models to produce marketing copy such as ads, landing pages, emails, and product descriptions. It is fast and genuinely useful for variants, compression, structure and repurposing. Its weakness is that a model predicts the most ordinary next word, so its default output drifts towards the category average rather than a distinctive voice.

Why does AI-written copy sound generic?

Because a language model is trained to predict the most likely continuation, and averaged across everything written on a subject, the most likely continuation is the most ordinary one. A voice is a consistent deliberate deviation from that average. The model is not failing to copy your voice; it is doing exactly what it was built to do, which is regress towards the mean.

How do you make AI write in your brand voice?

Stop describing the voice and start showing it. Adjectives like “friendly but professional” describe a huge region of language and narrow nothing. Six to ten paragraphs of your own best writing steer far harder, and a banned list of specific constructions you would never use narrows harder still. Examples do the work that descriptions only gesture at.

What should a brand voice document contain?

Five things: samples of your own best writing chosen for sound rather than performance; a banned list written as actual phrases rather than vague warnings; rhythm rules a model can check itself against, such as maximum sentence length; three before-and-after pairs showing a flat version and your rewrite; and one sentence naming your stance towards the reader. Short enough to paste into any tool, specific enough to enforce.

Which parts of copy should stay human?

The opening line, the central argument, and the close. Those carry the persuasion and are where averaging does the most damage, since a generic opening is indistinguishable from every other one the reader has already skipped. Let the model handle variants, compression, structure, repurposing, and the blank page.

Is AI copywriting bad for conversions?

It depends entirely on what it replaced. Used for volume work around a human-set position, structure and key sentences, it removes a bottleneck without cost. Used to produce the whole asset unsupervised, it tends to produce copy that describes the category rather than the company, which competes poorly precisely where differentiation decides the outcome.