When you write with AI, who decides which parts of the story stay yours — you, or the tool's convenience?

This explores how a writer working with AI decides which parts of a story to keep and which to hand off. The corpus has no study of authors making that call on purpose, so what follows is evidence about where the line tends to fall and where it probably should. The main finding is that the line is usually drawn by accident, not by judgment.

Framing and convenience seem to set the boundary. Writers told they own the final product leaned on AI suggestions significantly more, while writers framed as composing their own work put their effort into revising it themselves. That happened regardless of how good the AI was (Does ownership framing change how much writers rely on AI?). When writers did take AI text, they edited only 23% of paragraphs, and the edits kept about 96% of the original (Do writers actually edit AI-generated text before publishing?). Writers even chose the AI version of their own paragraph 63% of the time (Do writers actually prefer AI-edited versions of their own text?).

That last result shows why gut feeling is a poor guide to what to keep. Writers liked the AI rewrites and also objected to the systematic distortions those rewrites introduced. Polish and distortion turned out to be entangled in the model, so "does this read better to me?" can't tell you what is safe to delegate (Can user preference guide AI writing tool alignment?). A better rule has to come from what AI does to stories.

The evidence points at structure. Across five major LLMs, AI fiction over-explains its themes, favors tidy single-track plots, and avoids moral ambiguity. Human stories use nonlinear time and leave things unresolved (Do AI stories explain their themes more than human stories do?). A classifier separated AI from human fiction 93.2% of the time using only discourse-level features such as character agency and chronological order. It kept 97% of its accuracy with style cues removed. These choices resist humanization because fixing them takes a rewrite, not a surface edit (Can AI stories be detected without analyzing writing style?). So the plot's architecture (what happens in what order, who acts, what stays unsaid) can't be repaired after the fact and is the part worth claiming before drafting. Sentence-level polish is cheap to delegate and cheap to fix. My inference is that this fits a finding that GPT-3 places event boundaries closer to the averaged human consensus than individual people do (Do language models segment events like human consensus does?). A model's instinct for how a story breaks into events is the average reader's, which is the opposite of a distinctive structure.

The writing stage doesn't seem to be the line. In one study, writers used AI most for ideation, then for organizing thoughts, then for drafting, and they went back to ideation when blocked. Unexpected outputs sometimes opened new directions (How do writers use AI through different creative stages?). A model's surprises can feed early thinking as long as the author chooses what to keep. Two other things are hard to delegate. AI text is written for the prompter, not for an audience it has modeled, so who the story is for stays with the author (Does AI writing collapse the author-to-public relationship?). AI writing also lacks the internal appeal to a reader's attention that human writing performs (Does AI writing lack the internal appeal to attention that humans use?). Character work is more open. LLMs predict a character's choices better when given a persona profile plus memories relevant to that character's psychology (Can LLMs predict character choices from narrative context?). That suggests AI could act as a consistency check on characters an author has already decided on, not as the source of their choices.