AI-written stories can read smoothly line by line, yet feel flat: why do tension, character and meaning fall short?

This explores why LLM stories feel different beyond the sentence level: in tension, time, character and moral meaning, not fluency. The collection has no study of story generation itself. What follows is assembled from neighboring findings about how models generate text, and the story-specific links are my reading between them.

The main mechanism is that generation is smooth. Next-token prediction trains a model to continue toward its training distribution, not to explore counterpositions, so claims multiply without producing a new perspective (Does LLM generation explore competing claims while producing text?). Stories run on the opposite: a complication, a character wanting the wrong thing, a scene that turns against the reader's expectation. A smooth continuer is pulled toward the most expected next beat. A related note says an LLM output is a draw from a prior, not an observation of anything (Should we treat LLM outputs as real empirical data?). Read that way, an LLM story is a sample from the average of stories, not one author's risky choice.

That pull also has an emotional direction. GPT-4 turns negative prompts into neutral-positive responses about 86% of the time, and positive prompts rarely go negative (Does emotional tone in prompts change what information LLMs provide?). That was measured on answers to questions, not fiction. But it fits a familiar pattern: dark premises get softened and endings lean toward resolution. Moral framing shows a similar skew. LLM arguments use about 22% more moral language than human ones while matching human sentiment (Do LLMs use moral language more than humans?). In a story, that plausibly shows up as narrators and characters who name the values at stake instead of leaving them implied.

Time is a further gap. ChatGPT reasons about cause better than about temporal order, because causal connectives are explicit and frequent in training text while temporal order is usually implicit (Why do LLMs handle causal reasoning better than temporal reasoning?). Human storytelling depends heavily on the implicit: what gets skipped, flashbacks, what the reader has to infer about when things happened. A model that is strong on stated cause and effect and weaker on implied sequence would likely produce tidy, explicitly motivated plots. The grammar findings point the same way: LLMs handle simple structures well and degrade as embedding gets deeper, which suggests surface heuristics more than structural rules (Does LLM grammatical performance decline with structural complexity?). A plot with subplots nested inside it is the same problem at a much larger scale.

Characters are where the corpus is most concrete. When the same persona prompt is run repeatedly, variance across runs matches or exceeds variance across different personas, so model uncertainty, not a stable character, drives the output (Why do LLM persona prompts produce inconsistent outputs across runs?). What helps is scaffolding. On the LIFECHOICE benchmark (1,462 decisions from 388 novels), models predicted character choices better when given expert-written persona profiles plus retrieved memories relevant to the character's psychology (Can LLMs predict character choices from narrative context?). Consistent character psychology has to be supplied to the model; it does not come for free. When several voices interact, LLM groups converge earlier and surface less unique information than human groups (Do language model groups mimic human group reasoning patterns?). In dialogue, that suggests characters drifting toward agreement instead of pulling apart. Underneath all of this, the same weights produce different registers depending on the prompt, and each register inherits the quirks of its training distribution (Why do LLMs produce such different writing in chat versus posts?). A story prompt selects a register, and its narrative habits come with it.