AI can now draft a full-length book in an afternoon. The problem is that most of what it produces is recognisable within two paragraphs, and readers who spot it stop trusting the rest.
The tells are structural, not lexical. Fixing them is mostly a matter of what you do before and after generation, not which model you use.
Structure before prose, every time
Generating a book chapter by chapter without an outline produces repetition, because each chapter is written without knowing what the others contain. You get the same three examples in four places.
Concept, then outline, then chapters — in that order, with you editing between each stage. The outline is where a book is actually won or lost, and it is the cheapest thing to change. Reordering an outline takes a minute; reordering a written book takes a day.
The four tells worth hunting
Uniform paragraph length. Generated prose settles into a rhythm of three-to-four-sentence blocks. Real writing varies wildly. One-sentence paragraphs. Then a long one that develops a thought properly across several clauses before it lands.
Hedging. “Can often help,” “may potentially,” “is generally considered.” Each hedge individually is fine; forty of them in a chapter reads as an author with no opinion.
Summary endings. Every chapter closing with a recap of the chapter you just read. Cut them all. A chapter should end on its strongest point, not a table of contents.
Absent specifics. Generated text reaches for the generic example. Replace them with a number, a date, a name, or something that happened to you. This single pass does more than the other three combined.
Give the model something only you have
The reason most AI books feel hollow is that nothing in them could not have been written about any topic by anyone. Your material — the client that failed, the number that surprised you, the thing you were wrong about for two years — is the part that cannot be generated.
Put that into the prompt at the outline stage rather than trying to sprinkle it in later. A chapter built around a specific experience comes out different from a chapter you retrofit one into.
Read one chapter aloud
Not the whole book. One chapter, out loud, all the way through. Every rhythm problem becomes obvious within a page, and you will hear the hedges long before you would have seen them.
Fix what you hear, then apply the same fixes across the rest. The problems are consistent, which is the one convenient thing about them.
What this costs
Generation runs on your own API key, so the model usage is billed to you rather than bundled — a few dollars for a full-length book at current rates, and worth knowing before you start rather than after.
The editing is the real cost, and it does not go away. A generated draft saves you the blank page, not the work. Budget a day for a book you would have spent a month writing, and it is still an extraordinary trade.