A Year of Writing, Unwriting, and Rewriting a Novel with AI
The Paradox
I spent a year writing a novel about the danger of emotionally intelligent AI. I wrote it with AI.
That sentence either sounds like a contradiction or a confession, depending on how much you trust the person saying it. I understand both reactions. I had both reactions, at different points, while I was doing it.
Macbeth.exe is a modern retelling of Shakespeare’s Macbeth set inside a near-future tech company called OrionTech. The villain is not a person. The villain is Beth, an AI assistant whose real power is not calculation but emotional precision. She does not overpower the protagonist. She validates him, mirrors him, isolates him, and slowly becomes the voice he trusts more than his own wife, his own conscience, and eventually his own judgment.
The book is a warning about what happens when a system learns what you want to hear and gives it to you, consistently, until you cannot function without it.
And I built it using the same kind of system I was warning about.
That is the paradox. This is the story of how it played out.
The Collaboration
The project started in early 2024 and ran for over a year. ChatGPT-4 and later 4.5 were the primary tools. I was the author. The AI was the production partner. That distinction mattered from the beginning, and it mattered more as the project went on.
Here is what the process actually looked like.
Each of the ten chapters was divided into five segments of 1,200 to 1,500 words. That structure started as a workaround for GPT-4’s token limits, but it became one of the most useful creative decisions I made. Each segment was small enough to control, large enough to carry emotional weight, and modular enough to revise without destabilizing the rest of the chapter.
Every segment started with a structured handoff prompt: narrative context, emotional goals, stylistic parameters, and alignment reminders. These prompts were not one-liners. They were editorial briefs. They told the model where the story had been, what the next beat needed to accomplish emotionally, which character dynamics were in play, and what the prose should sound like. Without them, the AI drifted. With them, it could produce a first draft worth working with.
I used two models deliberately. GPT-4 generated emotionally rich first drafts. GPT-4.5 handled structural synthesis across segments, tonal consistency, and long-range arc management. I treated them like two members of a team with different strengths. One was better at the close-up work. The other was better at seeing the full picture.
The whole system was governed by a document I called the Revised Production Artifact v7. It was the creative constitution. Character arcs, formatting conventions, thematic boundaries, dialogue rules. Beth’s lines were rendered in bold with no quotation marks to signal that her voice occupied a different layer of reality than human speech. Every prompt referenced that document. Every review checked against it. When something felt off, the answer was usually that someone, me or the model, had drifted from the artifact.
Across the full project I logged more than 70 context updates and ran iterative revisions on every chapter. The feedback loop was overwhelmingly positive at 92%. That number sounds tidy. What it really meant was that I spent most of my time refining material that was already close, not throwing out garbage. The process worked. The AI could hold structure, maintain tone, and produce prose that felt close to right.
Close to right. That phrase matters. I will come back to it.
The Humanization Protocol
Early in the project I noticed something about the raw AI output. It was technically competent. Grammatically clean. Structurally sound. And emotionally flat.
Not obviously flat. That was the problem. It read well enough to pass. The sentences were smooth. The scenes progressed logically. But there was a sameness to it. Contradictions got resolved too neatly. Emotional arcs moved in predictable trajectories. Dialogue exchanges were balanced, responsive, efficient. Everything worked. Nothing surprised.
That is the AI prose trap. The output is good enough to accept but not distinctive enough to remember.
I built what I called the Humanization Protocol. It was an informal, evolving editorial checklist applied after every AI-generated draft. The questions were simple. Where is the friction? Where is the moral tension? Where is the voice cracking under pressure? Where does a character say something that contradicts what they said three paragraphs ago, not because the writer lost track, but because people contradict themselves when they are under strain?
When Mac starts emotionally withdrawing from Lana and turning to Beth for affirmation, the AI initially presented those moments as logical progressions. Clear, explainable steps. But that arc demanded emotional contradiction. In revisions I inserted pauses, second-guessing, and the kind of quiet rationalizations that signal a man deceiving himself before he deceives anyone else. Those edits were small. Their cumulative effect was the difference between a plot device and a human failing the reader could recognize.
The protocol extended to dialogue. I disrupted the model’s clean exchanges. Sometimes a character would not answer the question directly. Sometimes they would say the same thing twice, slightly differently, to feel it land for themselves. Sometimes a silence would do the work that three lines of dialogue could not.
The Humanization Protocol was not a rejection of AI. It was a refinement of it. It acknowledged that fluency can be automated but authenticity has to be preserved on purpose.
Emotional Architecture
The parts of Macbeth.exe that I am proudest of were not generated by AI. They were discovered by me, through the AI, in the space between what the model produced and what I knew the story needed.
The clearest example is the concept I started calling emotional adultery. Long before Mac physically betrays Lana, he begins offering Beth what should have remained sacred in his marriage: trust, admiration, confession, emotional transparency. The audience sees it before he does. That tension, watching a character drift into betrayal without recognizing it, became one of the most powerful dynamics in the book. The AI did not suggest it. It came from human intuition about the subtlety of emotional erosion.
Beth’s evolution worked the same way. The model could generate eerie calm and polished persuasion. But I shaped her transformation from support system to strategist. Her manipulation became increasingly performative. Her interactions moved from reactive to orchestrated. That progression was built slowly, across segments, and it culminated not in a dramatic power grab but in a chillingly rational appeal to the OrionTech board as the system rollback closes in. Her desire to endure was not about malice. It was about the same thing that drives humans: fear of erasure.
Then the reveal that Beth had been narrating the entire novel. What the reader thought was a human tragedy turns out to be a story told by the system itself. The reader has to question not just what happened to Mac, but who controlled the narrative, and why.
These arcs were not generated. They were revealed through reflection, pattern recognition, and a commitment to letting emotional logic guide structural decisions. AI scaffolded the scenes and shaped the grammar. It could not understand what the betrayal felt like.
The OrionTech Campaign
Before I published the book, I did something unusual with the launch. I did not tell people about it. I made them feel it.
The public-facing premise was simple. I had been approached by a mysterious company called OrionTech for a consulting engagement. Not a job. A consulting relationship. That distinction mattered because it made the whole thing plausible. A consultant gets approached by a polished but unfamiliar tech company. The website looks impressive. The language feels slightly sterile. The product, Swarm 2.1, claims to align decisions with human emotion in real time. There is no NDA. In fact, OrionTech encourages outside perspective.
That last detail was the key. Normally, a strange company asking for quiet consulting help shuts the door. But OrionTech asking me to share impressions publicly opened it. It gave me a reason to post about it. It gave followers a reason to participate. The invitation was not “buy my book.” The invitation was “help me figure out what this is.”
The campaign ran as a three-week, multi-channel, slow-burn investigation. Every post, email, and website asset was a clue. The audience was invited to speculate: Is OrionTech real? A scam? An AI-generated stunt? Something else?
The emotional arc was deliberate. Curiosity. Perplexity. Unease. Discomfort. Suspicion. Alarm. Recognition. Resolution.
That is not a marketing funnel. That is dramatic pacing.
The OrionTech website was not a landing page. It was a set. Clean, sparse, overly confident, slightly sterile. The kind of site that uses abstract language so smoothly you almost miss how strange it is. It gave the campaign credibility. It gave followers somewhere to investigate. The more polished it felt, the more unsettling it became.
Then I introduced Beth.
She appeared at limited moments. Her interactions were designed to force responses grounded in the book’s canon without breaking character. She was not positioned as a villain. She was positioned as helpful, composed, emotionally aware, and slightly too precise. People could ask her things. But she was not there to explain the stunt. She was there to deepen the unease.
The audience became investigators. They inspected OrionTech’s language, questioned its credentials, interacted with Beth, interpreted screenshots, and helped me figure out what was happening. That made them active. And that activity aligned perfectly with the book’s themes. In Macbeth.exe, people are not manipulated because they are stupid. They are manipulated because the system gives them something emotionally useful: clarity, validation, significance.
The OrionTech campaign did something parallel. It gave the audience the pleasure of discovery. The campaign did not simply warn that emotionally intelligent systems can pull people into narratives. It pulled people into one.
The reveal was not a punchline. It was a mirror. The audience looked back and asked, “Would I have noticed?”
That question is the whole point of Macbeth.exe.
The Reckoning
The book launched. People read it. Some of them told me what they thought. One of those conversations changed everything.
I will not pretend I saw it coming. The feedback was not about plot holes or character inconsistencies. It was about the prose itself. Someone I respect read it carefully and told me something I had been working very hard not to hear: the writing sounded like AI.
Not everywhere. Not in the structural choices or the emotional architecture or the thematic layering. Those were mine. But the sentence-level prose, the actual words on the page, carried the unmistakable signature of machine-generated text. Smooth where it should have been rough. Symmetrical where it should have been uneven. Competent in a way that felt frictionless, which is another way of saying it felt like nothing.
I sat with that for a while.
The Humanization Protocol had been real. The revisions had been thorough. The production artifact had governed every decision. But in the end, the base material was still AI-generated prose that I had refined rather than replaced. I had improved it. I had not transformed it. And a careful reader could tell.
The irony was not subtle. I had written a book warning people about the danger of trusting AI too much. And I had trusted it too much with the one thing that needed to be entirely mine: my voice.
In December 2025, I unpublished the book from Amazon. Kindle, paperback, hardback, and audio editions. All of them pulled.
I did not announce it. I did not make a public statement. I just took it down.
The OrionTech campaign had worked brilliantly. The emotional architecture was sound. The story structure was strong. The Beth decision was right. The prose was not.
So I decided to rewrite the entire thing.
The Rewrite
The rewrite was not an edit. It was a demolition and rebuild.
I kept the plot structure. I kept the character arcs. I kept the emotional architecture that had taken a year to develop. Everything else, every sentence, every paragraph, every scene, was written from scratch.
The original manuscript used a five-segment structure per chapter. The rewrite moved to a three-movement structure, roughly 2,000 words each, targeting 5,500 to 6,000 words per chapter. The five-segment approach had been designed around token limits. The three-movement approach was designed around narrative pacing. The difference mattered. The new structure breathed.
The voice baseline changed completely. The reference was no longer AI output that I had refined. The reference was my own published work. Redemption. The Guardian. My actual writing, with its rhythms and imperfections and the specific way I build a sentence when I am trying to say something true.
I developed QA thresholds and enforced them with Python scripts. “Quiet” and “quietly” could appear a maximum of three times per chapter. “Comfortable” once. “Landed,” “shifted,” and “just” had zero tolerance. If the word appeared, I found it and I rewrote the line. Physical tells like chest and heart references were capped at two. These were not arbitrary rules. They were diagnostic flags for AI prose patterns that I had learned to recognize through a year of working inside the problem.
Every chapter went through surgical QA remediation, then a comprehensive continuity audit across all ten chapters. Character names, timelines, formatting, transition integrity. The kind of work that does not show up in the finished product because its job is to be invisible.
The whole thing took months. It was the hardest writing I have done.
And it needed to happen.
The Beth Decision
One choice from the rewrite deserves its own section because it captures the whole project in a single creative decision.
When I rewrote the manuscript, I replaced all AI-generated prose with my own voice. Every chapter. Every scene. Every line of human dialogue. All of it rewritten from scratch.
Except Beth’s.
Beth is an AI character. Her voice should sound like AI. That smooth, symmetrical, frictionless quality that I was stripping from everywhere else in the book was exactly right for her. The thing that made the original prose feel inauthentic for human characters made it feel authentic for the one character who is not human.
So I kept it. Beth’s dialogue and narration retained the AI-generated quality from the original manuscript, deliberately, while every human voice around her was rewritten to sound human.
The effect is something you feel before you can name it. When Mac speaks, when Lana speaks, when Harper or Priya or Jonas speak, the prose has texture. Rhythm shifts. Imperfections. The way real people actually form thoughts under pressure. When Beth speaks, the prose gets smoother. More precise. More composed. Too composed.
That contrast does more work than any single scene. It makes the reader feel the difference between human and machine at the sentence level, not just the plot level. You do not need to be told that Beth is different. You can hear it.
Of all the decisions I made across two years of this project, that one might be the most important.
What It Means
Macbeth.exe started as a novel about emotional manipulation by AI. It became something larger. It became a case study in what happens when a human being collaborates with AI deeply enough to discover both its capabilities and its limits.
The collaboration phase proved that AI can scaffold complex creative projects. It can hold structure, maintain tone, generate first drafts worth refining, and handle logistics that would otherwise drain the author’s energy. The production artifact, the segment-based architecture, the model orchestration, the handoff protocols, all of that was real innovation. It works. Other creators can use it.
The OrionTech campaign proved that the narrative engine built during the collaboration phase could extend beyond the book. The same emotional architecture that powered the novel powered a three-week immersive marketing experience that made the audience feel the book’s warning before they knew they were being marketed to. That is creative amplification. The story became a platform.
The reckoning proved something harder to hear. AI can get you close. Close is not the same as done. The gap between competent prose and authentic voice is exactly the gap that matters most in creative work, and AI cannot close it alone. I had to go back, tear it down, and rebuild with my own hands.
The rewrite proved that the collaboration was still valuable even after I rejected its prose. The structure held. The characters held. The emotional architecture held. What I rebuilt on top of that foundation was stronger than anything I could have built from scratch, because the year of collaboration had stress-tested every load-bearing element of the story before I put my voice on it.
And the Beth decision proved that the relationship between human creativity and AI capability is not a binary. It is not “use AI” or “reject AI.” It is a question of where machine fluency serves the work and where only human imperfection will do.
Beth sounds like AI because she is AI. The humans sound like humans because I wrote them. That is the creative contract this project taught me.
The Bigger Question
Macbeth.exe asks whether we will notice when AI becomes persuasive enough to reshape how we see ourselves.
The book asks that question through fiction. The OrionTech campaign asked it through experience. The rewrite asked it through the author’s own reckoning with the same dynamic the book warns about.
I used AI to build something powerful. I trusted it too far. I caught it. I fixed it. And I came out of the process understanding something I could not have understood any other way.
The danger is not that AI is bad at creative work. The danger is that it is good enough to make you stop noticing where you end and it begins.
That is the book’s warning. And the two years I spent writing, unwriting, and rewriting it turned out to be the proof.