Why does AI writing all sound the same?

Because a model writes from the average of everything it has read. Ask it for a paragraph about your business and you get the paragraph most likely to follow your question, which is the paragraph every other business would get too. It isn't wrong, it's typical, and typical is what a reader skips.

Can a tone of voice document fix AI writing?

A generic one can't. 'Friendly, professional, approachable' describes the average, so it gets you the average with the edges sanded off. One that works is closer to a person written down: the things you never say, the rulings you've made and when, the words you reach for and the ones you ban. The model can rewrite against that all day. What it can't do is supply it.

Is it dishonest to write with AI?

Not if the position is yours. The model can carry an argument, check a figure and catch a tell. It can't hold an opinion on your behalf, and it can't tell you what it was like to be there. Those have to come from you, and if they don't the reader can tell.

Fighting The Average - Tony Cooper

I asked an outside model to review a piece I’d written, and it asked me to soften the one line I meant most. It wasn’t wrong about English. It was doing what every model does when it has nothing else to go on, which is giving me the average.

The review

The piece was Semrush Has Had Its Day. In the middle of it sits a line I believe: I think ranking reports are for the insecure.

ChatGPT wanted it softened. It also called one of my claims too absolute, the claim that Search Console already had the positions my old rank tracker was paying to collect. So I measured it. Of the 255 phrases where the tracker found a page, Search Console already had 251.

The claim was 98% true, and the suggestion was to make it sound less sure than the evidence.

ChatGPT can’t see my tooling or my reasoning, so it doesn’t know I’ve got a voice document. It reviewed from the only place it could, which is the average of everything it has ever read, and the average hedges.

I keep fighting against the average.

A model arrives full

When I wrote software the old way, every hour went on adding: a new feature, a new class, a new integration. You started from an empty file, and the craft was knowing what to put in.

A model doesn’t start empty. It arrives with every business book, every agency blog post and every piece of advice it was trained on, all of it switched on and all of it plausible. Ask it about a busy month and it suggests hiring someone. Ask it about growth and it says scale. Ask it to write and it writes the sentence most likely to come next, which is the one everybody else got.

So the craft turns round. Working with AI, you subtract what you don’t need, because the toolset comes with everything bundled in.

Most of what I’ve built is taking things out

Westworld, the system that runs my business, looks like a lot of rules from the outside. Read them and most of them are things the model mustn’t do.

There’s a list of advice it isn’t allowed to give me, because a solo practice with a ceiling of 30 clients isn’t a company waiting to scale. There’s a voice document that’s mostly bans: no apologising, no telling the reader how to feel about the sentence they’ve just read, no “most people” positioning me above a crowd. There’s a rule that any figure carries the check that produced it, because an unchecked number is the average’s guess at what’s probably true.

They all fire where the average would otherwise leak in. One fired on 1 October. I was publishing an insight on my commercial site, and the first draft promised readers I’d come back with more numbers later. I’d never offered that. It’s what the average writer says at the end of a piece, so the model said it, and the voice check cut it before anybody read it.

Subtraction gets you clean, not alive

There’s a limit to all this, and I set it myself back in July: subtraction gets you clean, and only addition gets you alive.

Strip out every default and you’re left with something correct and empty. The tells are gone and so is anything worth reading, and that’s still the average, just with the obvious marks scrubbed off.

So the subtraction only clears the ground. The one thing you add back is direction, and it’s the one thing that doesn’t come in the box.

Heathcote Williams ended his 1970 play AC/DC on a world where every idea is available to everyone, and none of them has any force left. That was 56 years before anybody could ask a model for an opinion.

The weapon

People hate AI-written guff, and I accept that. It’s terrible, and it’s terrible for exactly the reason above: it’s the average, delivered fluently.

I’m not apologetic about using it, though.

AI writing is terrible, but writing using AI with a tone of voice is art. I believe that.

It’s my voice that makes it art, not the format. A tone of voice document that says “friendly, professional, approachable” describes the average and gets you the average back. Mine is closer to a person written down: 27 years of doing this, rulings with dates on them, the singers I’d rather sound like (Nick Cave, Elvis Costello, Jimmy LaFave, Damien Dempsey), and a correction added every time the model drifts. The model can rewrite against that all day. It can’t supply the conviction or the bit where you were actually there, which is why those lines stay mine.

Fighting the average comes down to two moves. Take out everything that isn’t yours, and then put in the one thing only you have.

The review

A model arrives full

Most of what I’ve built is taking things out

Subtraction gets you clean, not alive

The weapon

Read more about the practice: