Confession: I Used AI To Write My Book

Writing without AI in 2024 Is As Foolish As Creating Content on a Typewriter in 2022

It’s true. I used AI to write my book, Outsmart the Learning Curve.

Of course, if you use these tools with any depth, you know an AI chatbot couldn’t come close to writing a book like mine (as I write this in April 2024).

Perhaps, one could cobble together a few dozen prompts to generate 200 pages of text, but it wouldn’t be this book—it would be something far more disjointed, wordy, and clichéd.

And a lot of stuff would be made up.

Developing this book required human judgment, creativity, and critical thinking that today’s AI is just not up for such as: defining a unique premise (a success book inspired by ordinary people), interviewing dozens of people, and tying it all together in a sensible outline.

Then I had to actually write it, and I could use some help.

AI Saved Me Hours…On Tasks It Was Good At

I couldn’t depend on AI to actually write anything. Even when I told it to write in the style of my previous work, it spat out wordy, often generic, and sometimes over-excited text.

In addition, it consistently hallucinated non-existent research links. After a while I stopped even asking it about academic research supporting various techniques or strategies because I couldn’t trust it.¹

But for things it was good at, it saved me hours.

For example, in a passage about an art school applicant, I wanted to illustrate the Slade School of Art’s esteemed reputation. Name dropping a few famous alumni familiar to readers would be easy, right?

Wrong! A quick glance at the Slade School of Art Wikipedia page shows nearly 200 “Notable Alumni” dating back to the 19th century—none of whom looked familiar to me. Before AI, I might have spent hours combing through each artist’s work to see if they created something familiar to my readers.

Instead, I pasted the alumni list into an AI chat window and asked it to give me a few names who’ve created famous pieces that might be well-known to US readers.

A few seconds later the AI returned sculptor Anish Kapoor, who created Chicago’s Cloud Gate, and pop artist Peter Blake, famous for his Sgt. Pepper's Lonely Hearts Club Band album cover art. I double-checked the output (as you always should) and it was right!

Hours of research turned into a few minutes.

Anish Kapoor’s Cloud Gate

AI Is More Trustworthy with Data You Provide

I chose Anthropic’s Claude as my primary AI tool in part because it supports attaching large documents to analyze (up to 10MB for “pro” users at the time of this writing). I found that when you provide the content as input, the output is much more trustworthy versus depending on its general amalgamation of Wikipedia, Reddit, and the rest of the internet.

I frequently uploaded research in PDF form and started with “Please summarize” (why I say please is not clear). The summary often reworded the paper’s abstract (not that helpful) as well as mirrored the sometimes technical language of the study. But the initial results would give me enough so I could pepper it with questions to get what I wanted, in language more understandable to me (e.g., “Can you describe the procedure and results of experiment 1 in layman’s terms?”) In this context, AI was amazingly accurate and often far clearer than the original. In a minute, I could evaluate whether a 20 page research paper was worth reading.

However, being skeptical is essential when using AI tools. For example, I would often ask questions like “are there enough subjects in this study to be statistically significant?” When asked, the AI would give me honest answers, but never volunteer that in the initial summary—I suspect because study authors often sidestepped that issue if it was a weakness.

Treat AI as an Inexperienced Research Assistant

There’s plenty of advice about how to use chatbots in general and Claude specifically. My most valuable suggestion (outside giving it your own data to chew on) is to word questions as you would to a smart, but naïve college freshman.

• Ask questions rich in context. Brief questions don’t save you time, they yield spurious results. Like your naive freshman research assistant, AI chatbots almost always answer confidently, even if the question is poorly worded. It’s your job to phrase your question with all the context to get the best results.

• Don’t trust the initial output. Clarify and verify everything. As I write this, AI chatbots are reducing hallucination frequency and getting better at saying “I don’t know” but be skeptical.

• Don’t let its “feel good” opinions sway you. There’s a tendency for AIs to be overly positive about your biases called sycophancy. Essentially, AIs are programmed to please you. I just reread a conversation about optimizing the book’s subtitle. On each iteration, Claude leaned into whatever my last suggestion was and ended the response with “the latest version looks exceptional” or “no notes” or “this looks like our winner.” At first, the complements on your ideas are gratifying, but don’t let its praise fool you.

Collaborating Alone Is an Oxymoron

Even though I was careful about Claude’s opinion on the book’s subtitle, I was able to give it my entire ~200 page draft as context (yup), and we had a very intelligent conversation about it—in the vein of a discussion I might have had with my writing coach. The discussion often took me down paths I wouldn’t have considered myself. Of course, Claude doesn’t have the creativity or judgment of a human writing coach, but unlike her, Claude is always available and, unlike my wife, has infinite patience (love you, dear ❤️).

Throughout the writing process, Claude unlocked the incredible power of collaboration without the weight of involving another human being. Doing collaborative work alone sounds like an oxymoron, but it’s a remarkably powerful tool in our era of remote work. While I couldn’t have written this book without my human coach, my AI collaborator helped me write the book faster and undoubtedly made it better.

I would love to hear your tips and tricks for using AI in writing or other pursuits. Please use the Comments to share with everyone!

1. After the bulk of my writing was complete, I found a service called Consensus which shows promise for building trust in using AI for research.