Skip to content
Working with the tools, not about them

Writing With AI

Generated prose drifts towards the middle of everything it has read

The pull towards the average is not a stylistic flaw to be corrected with better adjectives; it is what the process does, and working with it requires structural choices rather than word choices.

By Adrian Novak3 min read

Editorial note. Independent reporting and analysis. Nothing here is sponsored or paid for. How we work.

Why the output has a house style at all

Ask several different tools for a paragraph on the same subject and a family resemblance shows up: measured, balanced, slightly formal, fond of tricolons, reliably hedged, structurally tidy. That resemblance is not a design decision anyone made. It is what the centre of a very large body of written English looks like.

Anything fitted to a broad corpus and then asked to produce a likely continuation will tend towards the typical. The typical is safe, and safety is what a general-purpose tool is tuned for, so the pull is reinforced rather than resisted. Distinctive writing is by definition unlikely writing.

This is worth understanding rather than merely complaining about, because it tells you which interventions can work. Asking for more personality operates on the same average, and what you get is the average version of personality, which has its own recognisable texture.

Adjectives do not move it far

Instructions like punchy, conversational, distinctive and bold produce a shift in register and not much else. They select a nearby region of the same distribution, and after the first paragraph the gravitational pull reasserts itself, which is why long generated pieces so often begin with an energetic opening and settle into the standard cadence by the third section.

The reason is that these words describe an effect rather than a mechanism. Punchy is what a reader feels; it is produced by specific choices about sentence length, verb strength, and how much context is withheld. Naming the effect asks for the result without specifying the means.

Instructions that name mechanisms fare better. Vary sentence length deliberately, including at least one sentence under six words per paragraph. Use no more than one metaphor per section. Lead with the concrete case and generalise afterwards. These are checkable, and checkable instructions hold up over a longer piece.

Constraint is the reliable lever, not encouragement

The most effective single technique is to forbid the specific habits that constitute the average. No opening line that restates the question. No triadic lists. Nothing in the final paragraph that repeats what the piece already said. Each prohibition removes a region of the distribution rather than asking politely for a different mood.

Supplying examples does the same job from the other direction, and does it more efficiently for anything with a house voice. A handful of samples of the writing you want moves output considerably further than a paragraph of description, because you have shown the target rather than described the neighbourhood.

Neither approach is permanent. Over a long generation the constraints weaken and the sameness returns, which is one of several good reasons to work in shorter pieces and assemble them yourself rather than requesting a finished long document.

The averaging is sometimes exactly what you want

For a policy notice, a set of instructions, a summary of a meeting or a routine reply, the average is the correct target. Distinctive prose in a safety procedure is a defect. Here the pull towards the typical is doing useful work, and fighting it wastes effort on something nobody will thank you for.

The honest split is between writing whose job is to be understood and writing whose job is to be worth reading. The first is served well by the middle of the distribution. The second is defined by its distance from it.

Knowing which one you are producing is most of the decision. A great deal of frustration comes from applying the standards of the second to a task that was really the first, or from shipping the first where the second was needed.

What sustained voice actually requires

A voice, in practice, is a set of consistent decisions: what gets left out, where the humour sits, how much the writer concedes, what they refuse to explain. Those decisions are yours to make, and once made they can be written down and enforced. They cannot be delegated, because delegating them returns the average.

The workable pattern for anything with a byline is to draft in your own voice and use the tool for the mechanical passes — repetition, structure, the sentence nobody will parse on first reading. The voice stays intact because it was never handed over.

And there is a slower effect worth naming. Reading a great deal of averaged prose shifts what sounds normal to you, and writers who work this way for months sometimes find their unassisted drafts have moved towards the middle as well. That is not a reason to avoid the tools. It is a reason to keep reading things they did not write.

Common questions

Can I train it on my own writing to fix this?

Supplying samples of your writing within the instruction moves output meaningfully towards your patterns and is the practical version of this. It captures surface features — rhythm, vocabulary, structure — rather than the judgements underneath them, so it produces a decent imitation of your voice on subjects you have already covered and a weaker one elsewhere.

Why do generated pieces so often end with a summary?

Because a closing recapitulation is overwhelmingly common in the explanatory writing that fills the corpus, so it is the most likely ending. It is also almost always deletable. If you find yourself removing it every time, put a line in the instruction and stop paying for it.

Is the sameness getting better or worse?

Both claims are made and neither is settled. Newer tools handle instructions about style more precisely, which helps. At the same time a growing share of published text has been produced or smoothed by these tools, which narrows the pool that future ones are fitted to. How that balances out is genuinely unknown.

Writing With AIvoicestyledraftingsameness
Adrian Novak
Deputy editor, Prompt After Prompt

Adrian has written about prompt craft, writing with ai, images & audio for most of the last decade and prefers a plain explanation to a clever one.