Why AI writing tools all sound the same

By Anwar Benhamada · August 6, 2026

Run the same brief through five AI writing tools and you get five versions of the same text. Same rhythm, same transitions, same slightly elevated register, same tidy three-item lists.

This isn’t a coincidence and it isn’t laziness on the vendors’ part. It’s structural, which means knowing the cause tells you which fixes are real.

Why they converge

Most sit on the same handful of base models. The interface differs; the thing generating text often doesn’t.

They’re optimised toward the average. These systems produce statistically likely continuations. Distinctive writing is, definitionally, unlikely — so the default output sits at the centre of the distribution.

Post-training pushes toward inoffensive. Models tuned to be helpful and harmless converge on a register that’s clear, hedged, and slightly formal — because that’s what avoids problems across every context at once.

Vendors add safety-net system prompts telling the model to be clear and professional. Reasonable individually; collectively they eliminate the variance that makes writing sound like a person.

The tells

Once you see these you can’t unsee them:

• Rule of three everywhere. Three adjectives, three bullets, three examples.
• Uniform sentence length. Human writing varies wildly. Model output has a suspiciously narrow distribution.
• Summary paragraphs that restate. Every section ends by telling you what it just said.
• Hedged claims. “Can often help to” instead of “does”.
• Vocabulary tics. Delve, tapestry, landscape, testament, crucial, robust.
• Perfect structure, no argument. Every point gets equal weight, because nothing is actually being argued.

That last one is the deepest. Good writing has a position and subordinates things to it. Model output is frequently well-organised and about nothing.

Fixes that don’t work

“Write in a unique voice.” Produces a generic idea of unique — usually more adjectives.

“Don’t sound like AI.” The model has no reliable access to what that means.

Turning up temperature. Adds randomness, not personality. You get unusual word choices in the same rhythm.

Chaining more AI tools. Passing output through a “humanizer” produces text that’s differently generic, and usually worse.

Fixes that do

Give it your actual writing. Three or four samples of prose you wrote teaches rhythm and vocabulary better than any instruction. This is by far the highest-return fix.

Specify a position, not a topic. Not “write about hosting costs” but “argue that advertised hosting prices are misleading and the renewal rate is the only number that matters”. A model with something to argue produces structure that serves the argument.

Ban the tells explicitly. “No summary paragraphs. Vary sentence length — some very short. No lists of three unless there are exactly three things.” Negative constraints work far better than positive style instructions.

Give it constraints that force choices. A word limit that’s genuinely tight forces subordination — which is what creates a point of view.

Use it for structure, write the prose yourself. Often the honest answer. Outlining, ordering an argument, and finding what’s missing are things it does well without touching voice.

When it’s the wrong tool

If the value of the writing is the voice — a personal newsletter, a founder’s post, anything where people read because of who wrote it — the tool is working against you.

For that work it’s still useful upstream: what am I missing, what’s the counter-argument, is this ordering right. Then write it yourself. The distinction is between using it to think and using it to type.

The test

Read the output aloud.

If it sounds like a competent stranger who has no stake in the subject, that’s what you have. Editing won’t fix it, because there’s no position underneath to sharpen — and you’ll spend longer repairing it than writing it.