I Gave AI More of the Sentences. My Writing Got More Complicated.

The interesting shift was not from human writing to machine writing. It was from local prose production to causality, counterfactuals, and testing an entire fictional world against itself. This article was written with AI. Substantial parts of the prose were AI-drafted under my direction and revised through the same process this article describes. The first draft was also an almost embarrassingly useful demonstration of why that sentence tells you very little about how the work was actually made. It had one central argument, then performed approximately twenty-five pirouettes around it. The same point appeared under several headings. Short sentences isolated themselves into dramatic little paragraphs. Rhetorical questions arrived in flocks. Lists appeared because apparently no piece of online writing is safe from becoming 'five things I learned about something.' The argument was there, but the writing kept orbiting it instead of going from A to B. So I ran the draft through a checking system I have built from months of exactly this kind of failure. It checks whether an argument actually moves, whether a paragraph changes anything, whether the evidence supports the precise claim, whether rhetorical questions are doing work, whether lists are real lists or merely formatting habits, and whether the conclusion is explaining the article I have just spent ten minutes reading. That system is now checking this article. Which makes the article not only an argument about AI writing, but also a reasonably well-documented patient. Bad AI prose is easy. It repeats itself, overexplains, and occasionally sounds like a motivational poster has been promoted to middle management. I can delete that. What worries me much more is a beautiful sentence carrying the wrong thought. AI is extremely good at making incomplete shapes look complete. Give it most of an argument and it can infer the remaining shape. Give it a familiar kind of fictional character and it can produce the scene that normally belongs to that character. Give it several research findings pointing roughly in one direction and it can supply the bridge between them. Often the bridge is elegant. Sometimes the bridge changes the argument. A source says that somethingmay contribute toan outcome. A few paragraphs later, the prose says itexplainsthe outcome. Nothing spectacular happened. There is no invented study, fake quotation or absurd factual hallucination. One small phrase shifted, and the entire claim changed. I think of this as a craft-level form ofepistemic drift: a small movement in wording, category or certainty that quietly alters the structure built on top of it. I did not invent the term. In 2026,Anders Søgaard, Nina Rajcic and Ava Elizabeth Scottusedepistemic driftmore broadly for belief updating that becomes detached from underlying truth in systems where human minds and language models influence each other. My writing problem is narrower, but recognizably related: the local sentence remains plausible while the relationship between this sentence and the previous one has changed. That is exactly why the error is dangerous. A false date is easy to catch.May contribute tobecomingexplainscan survive fact-checking because every noun in the sentence may still be perfectly real. What changed was the logical load the sentence is carrying. The same thing happens in fiction. If a character is described in a summary as afraid, while the original scenes contain fear mixed with shame, anger, history, knowledge and obligation, an AI working from the summary may produce a completely coherent continuation of the wrong person. Nothing in the scene will necessarily look broken. The character has simply drifted. This is why I have become increasingly obsessive about going back to original material. AI can find the source, compare it with ten others and locate every place where a particular problem appears. But retrieval is not evidence, similarity is not continuity, and somebody still has to notice when the bridge moved. There is already a serious body of thought arguing that generative AI changes authorship by moving some human work away from direct sentence production. Luciano Floridicalls one version of thisdistant writing: the author becomes more of a designer, using an LLM to generate narrative while retaining creative control through direction and iterative refinement.Tiegue Vieira Rodriguespushes the idea further with an 'Apt Curation' model in which human authorship can reside in higher-order work such as architecture, evaluation, refinement and synthesis rather than necessarily in originating every sentence.There is even experimental creative-writing work that separates story structure from prose production.The 2026TombWriterstudyhad experienced writers work at the level of characters, scenes and beats while prose could be generated later. The writers valued the system particularly for structural discovery rather than prose production. So if my argument were simply that AI can produce prose while the human works on structure, selection and judgment, I would be arriving at a conversation that is already underway. That is not the part I find most interesting anymore. When I started using generative AI seriously, I thought the exchange was obvious: the machine would save me work, therefore I would write less. What actually changed was the level at which I could work. For most of my writing life, several cognitive activities happened at once. I developed an idea while writing the sentence that expressed it. I discovered a character while writing the scene in which the character acted. I solved structural problems while moving through the manuscript. Thinking, inventing, composing and typing were entangled tightly enough that it was easy to call the whole bundlewritingand never ask which component was consuming which part of my attention. Once some of the local work could be externalized, I could move farther away from the immediate paragraph. That did not primarily mean asking an AI to give me five possible scenes so I could choose my favorite. Increasingly, I was doing something else: changing one assumption inside an already enormous fictional system and asking what else had to move. If a character learns something here, can he still plausibly behave the same way three books later? If I remove an event, which later motivations lose their foundation? If a new interpretation of an old event is correct, which previous scenes acquire a different meaning? Which characters have enough information to discover the same thing? If they do discover it, what do they now have to do? And sometimes the answer is extremely inconvenient. A plot I like may stop being available because the character I have built would not behave in the way the plot requires. At that point, the question is no longer, 'Can I write around this?' It is whether writing around it would falsify the character. That is where the work becomes closer to model testing than ordinary brainstorming. I introduce a change and look for the consequences. I try the counterfactual. I search for the place where the system breaks. One of the most consequential discoveries in my fictional world did not happen while I was drafting a scene. I was folding socks. My brain was not occupied with producing sentences, and a connection surfaced that I had never consciously made before: by the logic already present in the world, the angels had effectively tried to kill God. I had never formulated the event that way. Once I had, I could not simply decide that this was wonderfully dramatic and insert it into the books. I had to find out whether it was actually true of the world I had already created. That meant returning to material scattered across the fictional system and testing the consequences. Existing scenes might change meaning. Some characters could understand what had happened while others could not. Later motivations might become stronger. Actions that had previously appeared reasonable might become impossible. The interesting question was no longer whether I liked my new idea. It was whether the idea was already latent in what I had written and, if so, whether refusing its consequences would make the world less coherent. AI is extraordinarily useful for this kind of work. It can retrieve the relevant material, compare events separated by years of story time, trace possible consequences and help identify the place where a new assumption makes something else collapse. That is not quite whatTombWriterdescribes as structural discovery. I am not primarily excavating the next scene from simulated possibilities. I am interrogating a persistent fictional canon as though it were a causal model. The direction of the question changes. Not simply: What could happen? But: Given everything that is already true, what is still possible? That distinction matters enormously to me. TheRelational Possibility Ontologygrew out of precisely this change in cognitive work and has since been published. I did not set out to invent an ontology as an intellectual side project to a novel. The scaffold changed first. Because I could spend less attention on some kinds of local production and retrieval, I spent more time examining relations, possibilities, constraints, agency, motivation and what follows when one condition changes. Eventually those questions stopped being only questions about fictional characters. They became the RPO. I cannot turn that personal history into a general claim that AI makes writers more intelligent. It does nothing of the sort automatically. But it gives me a concrete counterexample to the simple substitution story in which every cognitive operation handed to AI necessarily represents a corresponding reduction in human thought. In my case, some offloading created room for a different kind of thinking. That distinction turns out to be increasingly important in the research too. A very recent study of AI-assisted academic writing separates cognitive offloading into four layers: surface work such as grammar and local expression, structural work such as organization, idea generation, and reasoning itself. In that classroom study, deeper idea and reasoning offloading were associated with weaker later independent higher-order performance, while more bounded AI support performed better on the later unaided task than open collaboration. The researchers are appropriately cautious about causal interpretation, but the distinction between layers is more important to me than any single effect size. Because 'using AI to think' is not one operation. Neither is 'letting AI do some of the work.' There is a major difference between offloading the retrieval of twenty relevant scenes and offloading the judgment about what those scenes mean. There is a difference between asking AI to generate a sentence expressing an argument I have already built and asking it what I should believe. There is a difference between using a model to enumerate consequences and allowing the first plausible consequence it suggests to become the architecture of the story. Another recent study calls one dangerous versionReactive Writing.In more than a thousand observed AI co-writing sessions, writers often began evaluating AI-generated ideas before completing their own ideation. The suggestions then seeded directions the writers elaborated, even while the writers continued to feel that they were fully in control. That is the failure mode I worry about far more than whether the final sentence technically originated in my fingers. If I ask the machine what my character should want and then choose among its answers, I may be outsourcing exactly the layer of cognition I most wanted to preserve. If I already know the character well enough to ask, 'Given this new fact, show me every established place where his behavior would now become inconsistent,' I am using the same machine very differently. The useful question is therefore not: How much writing can I offload? It is: Which layer of writing am I offloading, and what does that free me to do? That is the distinction I wish I had understood at the beginning. There is an obvious problem with moving farther away from individual sentences. The sentences do not become harmless merely because I am thinking about something more interesting. Epistemic drift still happens. Character drift still happens. A source that merely contains the right subject can still be mistaken for evidence supporting the exact claim. A plausible reconstruction can still quietly replace something that actually occurred in the original material. I could try to hold every known failure mode in my head while also thinking about causality across dozens of books. I would rather not. So I externalize them. Repeated failures become criteria in checking systems. If an argument keeps circling instead of moving forward, the system checks for that. If an interpretation quietly promotes itself to fact, it checks for that. The same happens when prose breaks into theatrical one-line paragraphs, rhetorical questions begin reproducing in captivity, a complex argument becomes a list for no reason, or a conclusion explains the article the reader has just finished. I do the same thing with fiction. Once a constraint matters, I want it out of my working memory and somewhere it can be checked repeatedly. A character fact, a continuity dependency, a source rule, a distinction between what somebody knows and what the reader knows: I do not want the continued existence of any of those things to depend on whether I happen to remember them on Tuesday afternoon. The purpose of the checklist is not to automate judgment away. It is to automate the repetitive application of judgments I have already made. That leaves more of my attention for the cases in which I do not yet know the answer. This is where I expect people to disagree with me. Perhapswritingmeans producing language. A writer writes words. A novelist who designs a world, develops the plot and creates the characters but hands the actual prose to somebody else may be the originator of the story, but the person producing the sentences is doing the writing. Under that definition, if an AI generates the sentences of a paragraph, part of the writing has plainly been transferred to the machine. I think that is a perfectly defensible definition. I also think generative AI has exposed why it may not be sufficient. Literary work has always contained activities that are not physical inscription: inventing, structuring, choosing, rejecting, developing characters, tracking causality, controlling information, testing whether an action follows from a motivation, and recognizing when one apparently minor fact changes the meaning of an entire plot. Those activities used to be bundled together with sentence production closely enough that we had little reason to distinguish them. The person having the idea, following its consequences and producing the prose was usually the same person. We called the bundlewriting. AI can now pull the bundle apart. Once that happens, the old word becomes unstable. I can spend an afternoon changing one assumption and testing its consequences across a fictional world spanning dozens of books without producing a single final sentence. Under the narrow definition, I have not spent the afternoon writing. Yet that afternoon may alter the books far more radically than a day in which I produce several thousand polished words. We can call the first activitystory development,worldbuilding,conceptual authorshiporpre-writing. None quite solves the problem.Pre-writingbecomes especially strange when the process continues throughout drafting and revision and determines what the finished work is allowed to become. I do not think the answer is simply to stretchwritinguntil the word means anything remotely connected with making a book. The more interesting possibility is that one word had been hiding several different kinds of work because, until recently, they usually arrived bundled inside the same human being. Generative AI made the practical distinction visible. Now we have to decide what to call the pieces. Some of this article was written by AI. That should not be difficult to say. AI generated substantial portions of the wording, and calling every sentence 'human-written' would make the phrase meaningless. Calling the article simply 'AI-written' would conceal something else. The question comes from my experience. The distinction between local sentence production and causal model testing comes from observing what changed in my own work. My use of epistemic drift comes from repeatedly finding tiny shifts capable of corrupting an argument or fictional system. The checking criteria came from failures I decided were unacceptable. The argument changed because I rejected earlier versions of it, including earlier versions of this article. So I currently prefer a process description to a purity label: Human-conceived, human-directed, AI-drafted, human-revised. That description does not settle the philosophical argument about authorship. It makes the disagreement easier to locate. If writing is fundamentally the production of language, then AI-generated prose represents a substantial transfer of writing away from the human author. If writing is also the work by which a world, argument or idea is constrained, tested and determined, then sentence generation is one component of writing rather than its entire definition. The more interesting question may not be how many words the machine produced. It may be which layer of the work the human refused to give away.