Two months ago I wrote about using Claude to help finish a memoir I’d been sitting on since 2017. I stand by everything I said there.
I also want to be clear about where I stand generally: I’m not against writers using artificial intelligence to help with their writing. I’m not going to accuse someone of posting slop just because I recognize the fingerprints.
What I’m against is using AI so you don’t have to do any writing — outsourcing the thinking, the voice, the judgment, and then putting your name on the output. That’s a different thing.
What I’m doing with this memoir is what I called it last time: filling in the frog DNA. The raw material — the specific wrong turns in the certain streets of Lima, the embarrassing incidents, the stories I’ve never told publicly — all of that is mine. Claude, and at various points Gemini and Chad the PT, helped me get that material off voice recordings and into a shape my alpha readers (fact checkers, memory correctors) and my human editor can work with.
But here’s what happens when you use any tool long enough: you start to recognize its habits.
Claude finds a good metaphor the way a golden retriever finds a tennis ball: once it’s in its mouth, you’re going to see it for the next six miles.
(That one’s from GPT, which I find funnier than I probably should.)
After enough sessions across multiple tools working through this manuscript, I started noticing the same images coming back. Not because they’re wrong exactly, but because they’re not mine. In a memoir, that distinction matters more than it does anywhere else.
A partial list of metaphors and analogies I’ve had to cut:
- The brochure vs. the destination
- The lighthouse vs. the harbor
- The scaffolding vs. the building
- The compass vs. the map
- The bridge vs. the chasm
- The orchestra vs. the soloist
- The toolkit vs. the craftsperson
And for some reason, everything seems to happen on a Tuesday. So much so that I had to stop myself from mentioning that day of the week when sharing about going to school early on 9/11 because historically, it actually was on a Tuesday morning.
The metaphors aren’t bad. Some of them are structurally elegant. The problem is they’re bland and not to mention, available— the thing a well-trained system reaches for because they work just often enough.
My missions memoir book has a specific irony here worth naming.
At one point I was working with Gemini on a passage about the gap between what people expect when they go to the mission field and what they actually find. Gemini suggested framing it as the difference between “the vision and the brochures.”
I typed back: the brochures!!!!! NO! That’s even MORE AI.
Gemini agreed. It called the brochures analogy “a generic corporate metaphor that AI uses to describe missions”. Lazy. It was right. I have since found some version of “the brochure” in my own manuscript in no fewer than three places, all of them inserted during AI drafting passes, none of them actually mine.
The subtler problem is structural.
The metaphors I can search and replace. The structural habits are harder to catch because they look like good writing until you read them three or four times.
The lesson shoehorn. When I combined two or three stories in the same chapter — or followed Gemini’s suggestion to group-related ones — the AI would find a lesson. Always. Not because the stories had one. Because AI assumes stories exist to deliver lessons, and if the lesson isn’t visible, it will construct one and write toward it, and repeat it to make sure the reader gets the point.
The tells are phrases like:
- I’ve thought about that many times since then.
- I’ve thought about that over the years.
- Looking back now, I can see that…
No. I’m thinking about it right now, in this draft, for the first time since uncovering it in old emails I hadn’t thought of in two decades! The AI built me a wise narrator who has spent twenty years quietly processing this or that experience. What I actually had was a voice memo.
The valedictorian conclusion.
Every chapter wanted to end with a moral to the story. Some might have already had one naturally that I already included, but not ALL of them! When I was closing out sections — including one about the institutional dynamics behind how missionaries get sent — the AI would append a final paragraph along the lines of “Ultimately, this teaches us that the journey of growth is the true reward.” I kept cutting the last paragraph off entirely and letting a chapter end where the actual story landed.
Another thing about this that annoyed me as a writer is my humor and style is to just drop the hint and it’s there for the reader if they catch it. Too often AI would spell out the punchline or the “lesson” for the reader. Ugh.
The abstract concept label.
Across the intro and transitional chapters, the text kept trying to zoom out and categorize periods of my life with neat two-word phrases. The one that kept appearing: the unglamorous middle. It showed up enough that I went back to Gemini and asked for alternatives. When I pushed it for specific chapter examples of where we’d found it, it admitted it had been working from generalizations about the manuscript rather than the actual text. It hadn’t tracked the chapter numbers. It was describing a pattern it recognized in general, not one it could locate precisely in my book.
That’s its own kind of tell.
The sentimental essay translation.
Chapter 17 includes a story from a park near where we live in Chorrillos, Peru. My daughter was there with another kid who challenged whether she was actually Peruvian. Her response was to say, without explanation, “My dad is Canadian” — as if this settled the matter. The kid accepted it. That’s the story. It’s funny and weird and doesn’t need unpacking. I was including the story for context and texture of that season of life and occurrences.
The AI’s version included: “There is something in that exchange about what a kid raised between cultures reaches for…”
No, Claude, it’s not.
It turned a funny, weird moment into a sociology paper. It over-explained the joke. I cut everything after the dialogue and let her logic stand on its own.
Chapter 11 had a passage about a community in the in the mountains of Peru where a missionary couple I knew treated me well — the kind of people who made it obvious they were glad you were there rather than glad you’d arrived to do something for them. The AI’s version: “They honored me in the way that people honor someone they have decided to treat as a gift rather than a resource, and I felt it.”
That’s leadership-seminar language. It sounds like a quote someone puts on a powerpoint slide. The actual thing to say is simpler and costs less: “They treated me like a friend instead of someone who showed up to do a job they needed done.”
Even that still sounds AI-ish. I don’t remember what phrasing I wound up landing on in the end.
The point isn’t that AI is bad at this. It’s that AI is trained on a lot of self-help books, inspirational blogs, and writing that wants to mean something. When you ask it to help with a memoir, it brings that entire gravitational pull with it. The metaphors, the lessons, the tidy resolutions, the abstract labels for messy seasons — all of it is trying to make your life legible to the largest possible audience. Even when I have given it the manuscripts of my other books to imitate my voice and tone, it’s still going to do what AI is trained to do, whether I like the result or not.
A memoir that reads like it could have been written by anyone is not a memoir.
My alpha readers and friends reading this work for me have been helping me figure out where the AI labeled the wreckage instead of describing it. The unglamorous middle is a label. The long stretch of nothing happening is the wreckage.
What are the AI-isms you catch most often — in your own writing or other people’s?
