How I Built a Novel Like a System

Everyone is talking about AI right now. I know a little about it. I have been building with machine intelligence since before most people called it AI: recommendation engines, telemetry, data pipelines, and now the current generation of language models. I know these tools by hand. I know what they are good at, and where they break.

I also just spent two years writing a 200,000-word hard science fiction novel, and I used those tools throughout the process. How AI can be used for creative work is one of the most contested questions of the moment, and much of the argument is happening between people who have not actually tried it on a problem of this scale. So this is a practitioner’s answer: how these systems can help build a book, and how they really cannot.

The story had been growing in me for years: an interstellar object that is not what it looks like, real astronomical anomalies that nobody has ever explained, and institutions full of intelligent people making individually rational decisions that compound into catastrophe.

I found much of the underlying research and brought those subjects together. I built the logic connecting them, the characters, the institutional machinery, and the larger arc of the book and series. The context and the meaning were mine.

Something strange also kept happening while I was writing. Several ideas I had already developed were later strengthened by new scientific research. Papers appeared that supported parts of the physical or observational framework after those concepts were already in the story. The book was not following those discoveries. In some cases, the discoveries were catching up with the direction in which the fiction had extrapolated.

I wanted the book to exist at the standard the idea deserved.

So the first thing I did, in 2025, was test the obvious question: could AI write it?

No. Not close.

I evaluated it the way I would evaluate any tool for a critical system. I gave it the concept, the outline, and the science. What came back looked like a book in the way scaffolding looks like a building. Continuity drifted within chapters. The physics became confident exactly where it should have been careful. The voice flattened into something nobody would read twice.

I ended up spending more time repairing generated prose than working on the book itself. For a novel with a locked chronology, thousands of interlocking constraints, and a requirement that the science withstand an expert reader, generation was not a shortcut. It was a detour.

So I wrote the book. Then I put the machines to work on what they were actually good at.

Research. Much of the material was already in my hands. AI helped me investigate it more deeply, follow references, compare competing interpretations, and stress-test connections against the published literature. The novel draws on vanishing-star research, the acceleration anomaly of ʻOumuamua, the evolving chemistry of 3I/ATLAS, Kerr black hole physics, 1950s photographic-plate archives, impact modeling, space weather, and AI reliability. Every source carries its evidentiary status: peer-reviewed, preprint, disputed, historical, or theoretical. The research process depended on knowing the difference.

Verification. Two hundred thousand words contain thousands of facts that must agree with one another: dates, distances, velocities, masses, day counts, time zones, technical measurements, and who knew what when. The machines could search the whole structure repeatedly and flag contradictions that might otherwise survive for months. They flagged. I ruled. Sometimes the flag was right. Sometimes I was.

Cold reads. Before human beta readers saw the book, I used fresh-context machine reads with no knowledge of my intentions. They encountered the manuscript more like first-time readers and reported confusion, pacing problems, repetition, and continuity breaks without access to the explanations in my head. I fixed what deserved fixing and left untouched what did not.

Language. I have written millions of words professionally: documents, editorials, presentations, strategy, and public communication across the companies I built. A novel is a different discipline. It is very long, and every sentence sits inside a larger system of voice, rhythm, character, and consequence. I used AI as a language and editorial instrument. It could identify awkward constructions, repeated habits, or unclear passages and propose alternatives. I decided which sentences carried the weight of the book.

There is one more connection worth making.

AI is one element of The Fourth Visitor, but it carries one of the book’s central questions: what happens when humans and institutions begin to mistake machine output for judgment?

Several AI systems appear in the novel, built by different cultures and institutions, and none of them rebels. They identify patterns, calculate probabilities, and pursue the objectives they were given.

The failures begin around them. Confidence becomes certainty. A local objective is mistaken for the larger good. Outputs become armor for decisions people already wanted to make. The danger is not that the machines form opinions. It is that the humans stop forming their own.

The book states its rule:

The machine sees the pattern first. The human decides what it means.

That rule also governed how the book was built.

The systems could surface connections, inconsistencies, structural patterns, and possible alternatives. They could not decide what mattered, what belonged, what the evidence meant, or what the book was trying to say. Every suggestion was accepted, amended, or rejected according to the story I was building.

The same boundary governs both the novel and the process behind it. AI can be an extraordinary instrument. The failure begins when assistance is mistaken for judgment.

That is how The Fourth Visitor was built.

The book is free. Every citation in Chapters 14 and 15 is real. Look them up. It was written so that you could.