Chapter 13: Your Neanderthal Now Has a Copilot

Chapter 13 · Part Four: The Machine That Makes Average

Back to the wall from Chapter 2. Twelve different brands, one sentence.

I told you they weren’t real. I didn’t tell you how I made them.

I described our category to a model and asked it for a homepage. Twelve times. I barely changed a word.

In Chapter 2 that sameness was a story about people. This chapter is about what happens when you hand those same people a machine that’s better at sameness than they are.


The fluency engine

Here’s the plain version of how a language model works, and it’s all you need for the rest of this part of the book.

A model is trained to produce the most probable next words, given everything before them. The most probable words are, by definition, the most common words. And the most common words in your category are the ones everybody in your category has already written, which is exactly what everybody else’s model is producing, from roughly the same prompt, on the same Tuesday.

That isn’t a flaw in the tool. That’s the design. It’s a fluency engine. It was built to produce text that goes down smoothly, and smooth is what familiar feels like from the inside.

Now put that next to Chapter 1.

Your brain was already biased toward the familiar. It has been for forty thousand years, and it reports familiarity to you as a feeling of correctness. We’ve handed that brain a tool whose single greatest skill is generating the familiar, instantly, at volume, for almost nothing.

These don’t add. They compound. The model hands you the most probable sentence, your Neanderthal recognizes it and reports that it’s right, and the review round sands off anything the model happened to get interesting. Three filters, all tuned to the same frequency.


This has been measured

In 2024, Anil Doshi and Oliver Hauser published a study in Science Advances that I think every marketer using AI should read.

They had people write short stories. Some wrote with AI help and some without. When independent readers rated the stories one at a time, the AI-assisted stories came out ahead. More creative, better written, more enjoyable, and the lift was biggest for the writers who had started out least creative.

Then they compared the stories to each other. The AI-assisted stories were significantly more similar to one another than the stories written without help.

Every individual got better. The group got more alike.

That’s the wall of homepages, in a peer-reviewed journal. Every one of those twelve pages is good. Read one by itself and it’s clear, confident, well structured. Read all twelve and they’re one website.

This is the trap, and it’s why this chapter matters more than anything in Part Three. The review that would catch the problem looks at one piece at a time. One piece at a time, the AI version wins. The cost only shows up when you put your work next to everybody else’s, which is the only way your buyer ever sees it.


Belief two

So here’s the second of the seven beliefs, arriving where it’s earned.

AI didn’t make you faster. It made everyone faster at the same thing.

The gain is real. I’m not going to pretend otherwise in a book about honesty. A team using AI ships more, ships it sooner, and the average quality of each piece goes up.

But speed toward the middle isn’t an advantage when everybody else has the same speed and the same middle. It’s a treadmill. You produce more of what your competitors produce, faster, and the market experiences it as more noise from a category that already had plenty.


The test you should run

Before the keynote I asked myself whether the model was only converging because I gave it a boring prompt. So here’s a test you can run in ten minutes, and I’d rather you run it than take my word for anything.

Open whatever model your team uses. Ask it, twelve separate times in fresh sessions, for the most distinctive homepage it can write for a company in your category. Not a homepage. The most distinctive one.

Put the twelve heroes in a grid.

If they converge, you’ve learned the most important thing in this chapter: the machine’s idea of distinctive is the category mean. Asking it to be different doesn’t make it different. It makes it reach for the most common way of sounding different, which is its own kind of sameness.

If they don’t converge, you have a better model or a weirder category than I do, and you should tell me.


The anti-fluency toolkit

None of this is an argument against using AI. I use it more than almost anybody I know, and Chapter 15 is the inventory. It’s an argument for using it on purpose, against its own grain. Six techniques, in roughly the order I’d adopt them.

1. Locate the mean first, on purpose. Use the model to find out exactly what the obvious answer is, then treat that answer as the one thing you’re not allowed to say. This is the single best use of a fluency engine: it’s the fastest way ever invented to find the middle of your category so you can leave it. Ask it for the most typical version. Write it down. Ban it.

2. Inject a constraint. Every constraint from Chapter 7 works on a model, and works better than on people, because a model doesn’t resent you. Two words. No product nouns. No comparatives. Explain it to a plumber. A constraint makes the most probable answer illegal, and the model has to search further down the list.

3. Keep a forbidden-words list. Every category has twenty words that carry no information because everybody uses them. Write yours down and give the list to the model at the start of every session. In our corner of the market, the list starts itself: agentic, machine speed, seamless, empower, the work that matters. A word on that list in a draft is a word you didn’t choose.

4. Ground it in material the model hasn’t seen. A model can only remix what it has read, and it has read everything public in your category. It hasn’t read your customer calls, your sales team’s loss notes, the transcript of your CEO arguing with a prospect, or the strange thing a customer said on a podcast. Primary material is the only input that produces output nobody else can produce, because nobody else has the input.

5. Run an adversarial pass. After the draft, start a fresh session and ask a model to tear it apart as a skeptical buyer who has read eleven competitor pages this morning. What’s generic? What would a competitor say word for word? Which sentence could be deleted without anybody noticing? A critic that never gets tired is one of the most underrated uses of the tool.

6. Use a voice profile, and build it properly. A model with no instructions writes like the average of the internet. A model with a real profile of how a specific person writes, built from their own unedited writing and their corrections to drafts, writes like that person on a reasonable day. Chapter 15 covers how we build them, and the one thing that stops a voice profile from producing parody.

Notice what all six have in common. Each one adds friction back into a process that has been engineered to remove it. That’s the whole idea of this book, applied to a machine.


The Comfort Check

Chapter 13: writing with AI

1. The comfortable assumption. If each piece is better, the work is better.

2. The true objective. Make work that a buyer can tell apart from eleven competitors who have the same tools.

3. The redesigned approach. Use the model to find the category mean and ban it, write under a constraint, ground every draft in primary material nobody else has, and judge the output next to the competition instead of on its own.