I complained about AI slop. Then I built a tool that rewrites AI text.

Four weeks ago I published a piece about AI-generated posters taking over French supermarkets, and how every one of them looks like every other one. Today I am shipping a product that takes text a model wrote and rewrites it so it stops reading like a model wrote it.

If you want to call that a contradiction, go ahead, it is the first thing I would have said too. So let me put the argument on the table before someone builds a worse version of it for me.

What I actually said about the posters

The complaint was never that a machine made the image. Here is the line the whole article turned on:

The line I draw is not about the tool. It is about who is holding the point of view. AI inside a point of view is fine, and often great. A brand with a real visual identity, a real palette, a real way of framing its world, using a model to execute that identity faster, is using a tool. A brand with no point of view, using a model to supply one, is not using a tool. It is outsourcing the only thing that made it distinguishable from the shop across the street.

Carrefour's problem was not Gemini. It was that the output carried nobody's identity, so every chain ended up with the same soft studio light and the same glossy tomato. A generative model converges on the average of what it read. Ask it for a voice and it hands you the middle.

Text does the same thing, and it is worse in text because nothing warns you. A wrong-looking hand in an image is obvious. A paragraph that sounds like everyone is invisible until you have published three hundred of them and nobody can tell your blog from a content farm.

So the tool I wanted was not one that writes for you. It was one that puts the point of view back.

Where it came from

I keep a list.

It started as a note in a file and it now runs to about forty entries. Genuine. Genuinely. Truly. Deeply. Crucial. Nuanced. Delve. Robust. Testament. Landscape as a metaphor. Headings that describe the shape of the article instead of saying anything. Sentences that stop to comment on the writing they are part of. Perfectly balanced lists where every item opens with the same connector. Em dashes everywhere, which is how I got into the habit of banning things in the first place.

None of those words are wrong. They are just the ones that pile up when nobody is choosing, and once you have noticed them you cannot stop noticing them. Same problem as the posters: you spot them from three kilometres away.

Youmanize is that list, turned into a product, and then taken considerably further than a list can go.

What it does

You paste in text a model produced, and it comes back rewritten. Four tones out of the box, Natural, Professional, Casual and Concise, plus a strength dial that runs from a light pass to a full rework.

The one I care about is My Voice. You feed it things you have already written, it learns how you actually write, and it rewrites toward that instead of toward a generic idea of good prose. That is the whole thesis of the poster article, applied to text. Not "make this sound human", which just means make it sound like the average human. Make it sound like the specific person whose name is on it.

Everything factual survives the trip. Numbers, dates, names, links, code blocks, markdown, HTML structure, SEO keywords. The rewrite works on the prose and leaves the payload alone, which is the line between an editor and a generator.

There is a Chrome extension so it works wherever you already write, and a WordPress plugin that runs inside Gutenberg and the Classic editor, because that is where I live and it would have been strange to build anything else first.

What it does not do

This part matters more to me than the feature list.

Youmanize does not promise to make your text undetectable, and it never will. Text that a model wrote is text that a model wrote. Passing it through another model does not change that fact, it changes how it reads.

We say it on the pricing page in exactly these words:

We do not promise detector scores, and you should distrust anyone who does. Detectors are unreliable in both directions.

They are. They flag people who write plainly, and they clear machine output that has been shuffled around a bit. Building a business on beating them would mean building on a number that means nothing and moves every time a vendor ships an update.

There is a second reason, and it is not going away. Under the EU AI Act, machine-generated content can carry an invisible, machine-readable provenance mark. We make no claim to strip those, and we make no claim to remove the statistical fingerprints a model leaves in its word choice. Anyone selling you that is selling you a number they cannot control.

So, plainly: your text will read better and sound more like you, and it will still have been written by a machine.

Unless a person writes it

Which is why the top tier is not a model at all.

Pay for human review and an approved editor rewrites your text by hand. No model touches the final version, so there is no machine mark to argue about, because there is no machine. It starts at fifteen dollars per thousand words and comes back within twenty-four hours.

I like this tier for a reason that has nothing to do with detection. It is the only part of the product where somebody gets paid to write. Everything else in this industry is pointed the other way.

And it is open on both sides: if you write well, you can sign up as a reviewer and earn from it. When I wrote about Milo I said that an ecosystem where other people make money is the only kind I have ever seen win. I meant it there and I mean it here. A business that makes writing worth money again is a better business than one that makes writing worth nothing faster.

The part I am not going to pretend about

Some people will use this to hand in an essay, or to pass off machine output as their own work to someone who asked for theirs. I know that. Every editing tool ever made has had that problem, and saying so in the marketing copy would not stop a single one of them.

What I can control is what the product claims. It does not claim your text is human. It does not claim a detector will clear it. It claims the text will read like you, and if you want writing that came from a person, there is a button for that and it costs money, because a person is doing it.

That is a narrower promise than most of this category makes. I think it is the only one that survives contact with how these systems actually work.

Where it is

youmanize.me. Free tier at five thousand words a month, paid plans from nine dollars, human review from fifteen dollars per thousand words. Chrome extension, WordPress plugin, REST API.

I still think the supermarket posters are a problem. I still think a web where everything converges on the middle is a worse web. The difference between complaining about that and doing something about it turned out to start with a list of forty words.