I use AI every working day, including inside my drafting process, and I regard every published word as mine to answer for. If my name is on the book, I cannot send a dissatisfied reader to the chatbot that suggested the offending paragraph.
The help I ask for depends on where I am in the work. I might dictate what happens next and ask AI to turn that rough material into a first prose pass, or ask for alternatives to a transition I have already rewritten several times. Other passages I write myself, although any of them may go through an AI critique later. However the wording takes shape, I have to compare it with what I intended before deciding whether to keep it. A generated paragraph receives no special consideration for having appeared quickly.
That is the working relationship I want to examine here. I enjoy these tools enough to use them extensively, and the disagreement often begins at the point where their suggestion sounds entirely reasonable.
I give the model a job I can inspect.
Ghostwriting has made me attentive to the different kinds of work a book requires. Checking a character's age against an earlier chapter gives me a fairly definite question to answer; deciding how she responds to betrayal requires a judgement about this particular woman and the book around her. I need to understand that difference before I decide what help to ask for, because a model can offer an equally confident answer to either question.
As a critical reader, it is most useful when it directs me to a passage I can examine. I may ask it to find a gap in the logic or an explanation the text expects the reader to accept without enough preparation. Even with those instructions, I have to check whether the objection holds. Deliberate ambiguity can look like missing information, and a character's irrational choice may be exactly what her fear would produce. A dutiful attempt to repair either could damage the scene.
Repetition gives me another useful task because I can check the model's findings against the manuscript. If it claims that I have used a phrase forty times, a search will tell me whether its count is right. Deciding what to remove takes more thought: an accidental tic and a phrase that a character deliberately repeats may look much the same in a list. For continuity checks, I likewise want the passages that establish the alleged contradiction, so I can see whether the model has missed an explanation elsewhere.
Drafting from dictation demands closer attention. The prose pass must preserve what I supplied while fitting the point of view and the voice around it, which gives me several ways to assess a sentence beyond whether it reads smoothly. If the wording changes the character's intention, I have acquired a new problem alongside my lovely paragraph.
I need that ability to inspect the contribution. When I cannot explain what would make the answer useful, I take it as a sign that I need to define the job more carefully.
A fluent improvement can damage the scene.
Imagine a romance in which the hero has a dangerous reputation and the heroine has good reasons to distrust him. In this hypothetical revision, I ask AI to clarify their dialogue, and it obliges by having him explain his motives and reassure her that she has misunderstood him. The exchange becomes easier to follow because much of its uncertainty has disappeared.
Whether that helps depends on what I need the encounter to do. He might have reached the point where an explanation is possible, or his refusal to explain might be the difficulty the relationship still has to overcome. Before accepting the revision, I would compare what the heroine believed before their conversation with what she can reasonably trust afterwards.
An editor has to distinguish between confusion that obstructs the story and uncertainty that sustains it. A repeated word may need removing while an uncomfortable silence needs protecting, even though both attract the same vague request to improve the passage. Once I know why the silence matters, I can ask for a clearer exchange that leaves it intact.
That is why a model's explanation of its own revision cannot settle the question for me. I have to test the suggested change against the scene's purpose.
I keep the verdict on my side of the desk.
When AI calls a chapter compelling, I want it to identify the choice or discovery that gives the chapter its force. If it can show how new information makes a character's next decision harder, I have a claim I can examine. Otherwise, I have a compliment from a machine I am paying to help me.
Criticism needs the same scrutiny. I can waste an afternoon repairing a problem that was never there if I treat confidence as a reason to obey, and the wasted effort becomes worse when the repair removes something the book needed. Before I change a passage, I want to be able to explain the problem in my own words.
For factual claims, I also need an appropriate source. An answer about a historical detail or the current market gives me a lead to investigate, and I have to follow that lead before I build a scene or an argument around it. Fluency will not tell me whether the claim survives checking.
That responsibility continues through publication. A contradiction that reaches the reader is one I missed; a revision that changes the narrator's voice is one I accepted. AI may help me catch either, but the decision to publish remains mine.
I describe the assistance accurately.
I am willing to say that AI contributes to parts of my drafting process, because pretending it only catches typos would misrepresent the work. The phrase “AI-assisted” covers such different practices that I would rather explain what the tool did and how I used the result. A continuity check and a generated prose pass involve different contributions, and readers should be able to understand which I mean.
They are also entitled to dislike that process. I can describe it honestly and stand behind the finished book without expecting everyone to approve of how it was made.
Here, that honesty includes showing revisions that failed and explaining why I rejected them. Invented examples will be identified as examples, while claims about readers or the market will need evidence beyond a model agreeing with me. I want this publication to be useful to writers who work with AI, and that requires enough detail for you to question my decisions as well as the model's.
When a suggestion sounds good until I test it against the book, the reason I reject it is worth examining. That moment of disagreement is often where the useful craft discussion begins.