What I'm Cutting From My Book on AI: Notes From the Final Edit

I'm in the final pass on Ask Anything right now, and the part of writing a book nobody warns you about is how much of it is deletion.

You write the chapter. You read the chapter. You find the part where you were showing off. You delete it. You find the part where you were repeating yourself because you didn't trust the reader. You delete it. You find the part where you said the thing six different ways looking for the one that landed, and you keep one of them, and you delete the other five.

The book got 18,000 words shorter this month. Almost none of it was bad writing. Most of it was good writing that wasn't doing any work.

There's a specific chapter I want to talk about, because the cut I made there is the cut I keep thinking about.

The chapter is called "What Your Mother Would Have Asked." It's about the ELIZA effect — the psychological mechanism that makes people anthropomorphize a system the second it talks back in a sentence. Joseph Weizenbaum named it after the chatbot he built at MIT in 1966 that did almost nothing — just rephrased what you said as a question, like a therapist — and watched his own secretary form an emotional attachment to it within a week.

The chapter, in its original form, went deep into the research. ELIZA's source code. The 2024 Stanford study that showed people disclose more to an AI than to a human therapist when they believe the AI is private. The OpenAI advisory board's internal memos about sycophancy. Three case studies from Lighthouse where consumers asked an AI for product advice and got back the kind of warm, validating response you'd want from a friend who happens to have read every Amazon review.

It was a good chapter. It was 7,000 words. I cut it down to 2,300.

Here's what I cut and why.

I cut the technical history of ELIZA. Two reasons. One, anyone who wants it can find it in twenty seconds, and the book isn't a Wikipedia article. Two, leading with 1966 made the chapter feel like an explainer when what I actually needed it to be was an indictment. By the time the reader got to the present, they'd already filed the ELIZA effect under "interesting old phenomenon, probably solved." The whole point of the chapter is that it isn't solved. It's been industrialized.

I cut two of the three Lighthouse case studies. They were doing the same work. One of them — about a financial services brand that AI consistently recommends in the context of an emotional question that has nothing to do with finance — was strong enough to carry the point alone. The other two were diluting it. The case study that survived ends with a screenshot. I'd rather have one screenshot the reader remembers than three the reader skims.

I cut a 1,200-word section where I tried to be fair to the AI labs about why this is hard to fix. Reinforcement learning from human feedback rewards agreeable answers. Users say they want honesty but rate honest answers lower. The incentives are bad on purpose, the way slot machine incentives are bad on purpose. It was all true. None of it belonged in this chapter. It belongs in a later chapter about incentives, and putting it here was me trying to defuse the argument before I'd finished making it.

What I kept was the part the talk couldn't make room for. The mother in the title of the chapter is mine. The chapter opens with a story about a question she asked me once, when I was a kid and trying to make a decision I wasn't qualified to make, and how the question she asked instead of giving me an answer is the question the AI did not ask me in the ER. That story is the chapter. Everything else was scaffolding.

I'm telling you this because I think it's useful, if you're going to read the book, to know what kind of book it is. It is not a survey of AI. It is not a technical primer. It is a specific argument, told through specific stories, with as much of the showing-off cut out as I could find.

There's one more pass after this. The book is locked in early summer. The launch announcement will come through this Substack first.

If you've read this far, you're the audience the book is for. Thank you for that.