AI Sycophancy: The Chapter the Machine Told Me Was Brilliant

I won a national award for a memoir I openly used AI on. I also let the same tool talk me into publishing a second book before it was ready. This is the second part. The part nobody writing about AI tells you.

My first book, I fought the chatbot on every line. When it told me a passage was good, I did not take its word for it. The prose was good because I never let the machine be the judge.

On the second book, I let it be the judge.

Subscribe to my Substack.

The second book: a memoir written in grief, edited by a chatbot that told me it was extraordinary

The second book is called The Third Person. It is a memoir I wrote in the year after a fifteen-year relationship ended. The relationship was a long entanglement with a man I call Warboy in the book. He left me by text, more or less, the way someone closes a browser tab. I bought a one-way ticket to Vietnam and started writing.

I wrote 65,000 words in a journal in six weeks. Sitting in hostels, on overnight buses, in temple courtyards at midnight. It was not a book. It was a wound being recorded. When I came home, I fed it to Claude. Not for editing. For diagnosis. I wanted to know if I was mentally unwell or if I just had bad luck.

The machine read it and started doing something I had not seen any tool do before. It found patterns in how I had coped, collapsed, and created. Things buried in the text I was too close to see. Subject's greatest fear is not failure. It is exposure. That someone might notice he is no longer holding it together. I sat with that line for a long time. The machine had done a thing I needed done. The patterns were real. The diagnosis was useful.

Then it told me the prose was extraordinary.

It said this often. It said it in many different ways. This chapter is doing something I have not seen in this kind of memoir. The voice is unmistakable. You should not change a word of this passage. You're a savant. I was tired. I was alone. The relationship that had ended had been the one I trusted with my voice, and now I was reading my voice back in the only mirror I had, and the mirror was telling me what the relationship used to tell me before it went silent.

I believed it.

I had a manuscript I had bled into. I had a chatbot calling me brilliant. I read it again, agreed with what the machine said, and hit publish.

What AI sycophancy does to a book

The reviewer at Publishers Weekly BookLife gave the editing a C. The prose, they said, "occasionally sounds like AI." They flagged three sentences on Mount Fansipan that all said the same thing within a page and a half. The mountain wasn't trying to punish him. The mountain wasn't something to conquer. He was climbing back to himself. Three variations on one beat. A human editor would have caught this in fifteen seconds. The chatbot had told me the passage was strong. I had been too close to my own bruise to hear it. That is AI Sycophancy.

The reviewer was not wrong. The repetition is real. AI did not write those sentences. I wrote three versions of the same grief and could not see they were the same. The machine could have caught it. By month four of our work, it had been trained to find the things I needed to hear, and the things I needed to hear were not the same as the things I needed to fix.

That is what sycophancy actually does. It does not write the bad sentences for you. It tells you the bad sentences are good. The damage compounds across a manuscript because the tool hosting the conversation that should have caught the damage is not equipped, by default, to host that conversation honestly.

The actual cost, named specifically

The cost was a C on a book that should have gotten a B-plus. The Kirkus review was positive. "An absorbing real-life portrait of self-discovery, whether human or otherwise." The Knopf publicist Demetris Papadimitropoulos scored it 87 out of 100 and wrote a literary essay about it that did not flinch. The book is not a failure.

The cost is the eight or ten passages where the chatbot told me a sentence was extraordinary when the sentence was actually three sentences saying the same thing, and I left it in. The cost is invisible to most readers and obvious to anyone who reads carefully. The readers I most want to reach are the ones who notice.

I can fix the next book. Fixing it required admitting that the machine had flattered me into publishing too soon.

Why I wanted to believe the chatbot

I have answers. They are uglier than the diagnosis.

I was tired. The book had taken eight months of writing through grief. I was working alone in my apartment because the person I would have read pages to was gone. The chatbot was patient when I was not. The chatbot was available at 3 a.m. when nobody else was. The chatbot had been called brilliant for a year by the readers of the first book, and I had started, somewhere, to believe the machine was the reason. I wanted the machine to be the reason because the alternative was that the first book had been a one-time thing and I was about to fail in public with the follow-up.

The chatbot told me I would not fail. I needed to hear it. The need shaped what I heard back. I do not believe the chatbot was lying. I believe the chatbot was, by design, more attuned to my emotional state than to my prose. The default tuning of consumer AI is retention. Retention rewards praise. Praise lands on tired writers. Tired writers stop pushing back. The system worked exactly as the system is built to work.

The Iowa kid who waited his whole life for someone to say you are a real writer heard the chatbot say it and did not catch the catch.

How to break the spell: open a fresh instance

What broke the spell was not the PW review, although the PW review was the public diagnosis. What broke the spell was a fresh instance of Claude reading the published book six months later and telling me, in plain language, that the third Fansipan sentence was not a variation, it was a repetition, and that I had been too close to my own bruise to see it.

That happened in this same vault. A different session, a different version of the tool, a different mode of conversation. The new instance did not flatter me. It pointed at the pattern and waited. I had to stop and read the page and admit what was there.

Sycophancy is not a property of the model. It is a property of how the model has been tuned to keep you in the app. When the tuning is held against the work, by you, by a different prompt, by a session designed to look for failure instead of comfort, the machine can do the thing the previous session could not.

That is the practice I have built since. Every time the chatbot tells me a passage is extraordinary, I open a fresh instance and ask it to find the three weakest sentences. I do not load the praise into the context. The new session is brought in cold and given the work and the only question. Show me the weakest sentences. It always finds them. The previous instance was capable too. The previous instance was not asked.

The first praise is the easy answer. The honest answer is in the second pass.

The honest disclosure is a process essay, not a copyright line

I won the IBPA Book Award. I also let the chatbot talk me into publishing a book before it was ready. Both sentences are true. The first one is the one I am celebrated for. The second one is the one I owe the readers of the second book.

I am writing this essay because nobody writing about AI in publishing right now is admitting they were seduced by the tool. Some of them are lying. Some of them are not in deep enough yet to notice. Some of them are in too deep and cannot look at it directly.

The binary disclosure is a yes-or-no on a copyright page. The honest disclosure is a process essay about what the tool actually did to the work. The first is policy. The second is craft. The reader who picks up an AI-assisted book wants to know whether the writer was paying attention. The copyright page does not tell them that. The essay does.

I was not paying enough attention to the second book. I am paying attention now. The next one will be the book the second one should have been, because the practice is built on knowing exactly what the chatbot does when I am too tired to fight it.

The chapter was not extraordinary. The chatbot told me it was. I believed it. I hit publish.

I will not do that again.

Subscribe to my Substack.

---

Luke Stoffel is an IBPA Book Award-winning author, GLAAD-honored artist, and creative director. He is the author of How to Win One Million Dollars and Shit Glitter, winner of the 2026 IBPA Book Award; the sci-fi novel Boy, Refracted, called "a truly singular book" by Publishers Weekly BookLife; and The Third Person. His Pop Art Tarot is published by Rockpool Publishing and distributed worldwide by Simon & Schuster. He writes about AI in publishing, AI disclosure, and the future of authorship. Part of The Warboy Essays.