AI stories repeat themselves in a way that most casual readers would never notice. Ask a chatbot to invent a piece of fiction and it may take you to a lighthouse, introduce a keeper called Elias, or pull another character from a surprisingly small collection of names. Cornell researchers examined 20,000 stories generated by four language models and found that certain words appeared far more often than the variety of the prompts would suggest, according to Cornell's explanation of the findings.
The result is funny enough to sound like a running joke, but it asks a serious question about artificial creativity. A machine can assemble fluent sentences and still return to the same narrative territory. The research describes a measured pattern in a set of models and prompts, not a rule that every chatbot will always write about lighthouses.
AI stories repeat even when the assignment changes
The research team used five prompts across four contemporary language models and collected 20,000 fictional outputs. In their analysis, eleven recurring words appeared in 88.3 per cent of the stories. More than half featured a lighthouse, according to Cornell’s account of the findings. The recurring vocabulary included names such as Elias, Mara and Elara, as well as professions including clockmaker and librarian.
None of these words is inherently strange in fiction. A lighthouse is a wonderfully atmospheric place to begin a story, and a mysterious keeper is a familiar character. The surprise is that a device sold on its capacity to generate endless alternatives can keep choosing similar ingredients when different users ask for something new.
The finding concerns frequency and diversity, not whether an individual story is emotionally effective. A skilled writer can make a familiar setting feel fresh. But if the same settings keep appearing without anyone requesting them, the range of the underlying creative process deserves scrutiny.
The lighthouse may be a fingerprint of training
Researchers Sil Hamilton and David Mimno looked for explanations in the material used to train and refine language models. Their research paper argues that the repeated words may be linked to preference data, examples and feedback used to encourage an assistant to produce answers judged helpful or acceptable. They say those words do not commonly occur at comparable levels in published literature or ordinary pre-training text.
That is a hypothesis drawn from evidence about data patterns, not a confirmed reconstruction of every company’s training pipeline. Commercial model builders rarely disclose every example or post-training choice. The researchers therefore cannot establish that all four systems copied from precisely the same dataset.
The general mechanism is understandable. If many pupils practise writing using the same model answers, the pupils may begin to reach for similar phrases. A language model also learns what a desirable response tends to look like. When the examples reward certain styles, characters or situations, those choices can be repeated long after the original examples have disappeared from view.
Good sentences can disguise a narrow imagination
Modern generators can vary sentence rhythm, include dialogue and assemble convincing descriptions. Those capabilities make repetition harder to spot. Two tales may use different verbs, weather and endings yet still feature the same isolated building, solitary caretaker and quietly unsettling revelation.
For readers who occasionally ask an assistant for a bedtime story, this may be no great problem. For a publisher generating large amounts of fiction, a school using AI to support creative work or a games company building characters, the problem becomes more practical. A supposed production shortcut could gradually make every output feel as though it came from one writer with a handful of favourite ideas.
A useful test is not simply whether a model avoids plagiarism. It is whether it can respond to a particular creative brief with enough independent alternatives for an editor to make meaningful choices. LiveAIWire has previously explored the challenge of how artists’ influence appears in AI images, where apparently novel material may still reflect patterns learned from other people’s work.
Why prompts are part of the answer
There is a tempting response to this kind of research: tell the chatbot not to mention lighthouses, and the problem is solved. Such an instruction may help with one visible symptom. It does not necessarily produce a wider repertoire of characters, conflicts, tones or settings.
A more informative experiment changes the shape of the assignment. Instead of merely requesting a surprising story, a user can specify an unusual everyday location, an emotional conflict, the length of the scene and a limit on familiar fantasy devices. Then ask for genuinely different outlines before selecting one. Compare what the model supplies, rather than assuming that a larger number of drafts guarantees variety.
It also helps to include real observations or personal details that the model would not have invented unaided. Someone who remembers the smell of a local swimming pool or a particular conversation at a bus stop brings material that a generic prompt does not contain. That human specificity can lead the machine away from the smooth but repetitive average.
This is not evidence that every AI story is bad
The study establishes an unexpected pattern across a defined experimental set-up. It does not rate the literary quality of each story against human fiction, establish that every reader dislikes the repeated vocabulary or prove that every model version behaves the same way. New model training and a different range of prompts could produce different figures.
Cornell says the research was due to be presented at the October 2026 Conference on Empirical Methods in Natural Language Processing. The publicly accessible paper was initially posted as a preprint in May. Readers should distinguish the research findings from a broad prediction that machine-written books will inevitably sound identical.
Another problem is judging an output by its fluency. A story can sound persuasive or confident without being original, just as a chatbot can sound persuasive when its factual account is unreliable. LiveAIWire’s reporting on confident but unreliable chatbot answers explores the dangers of mistaking a pleasant answer for a checked one. The same habit matters in creative work, although originality demands a different kind of check.
Editors can look for a hidden house style
A human editor reviewing AI-assisted fiction should consider patterns across the entire collection. Are characters always described as quietly determined? Does an unexpected letter repeatedly arrive just before the ending? Does every revelation involve a forgotten memory, an old book or a light in the distance? These similarities may matter even when no passage is literally duplicated.
A practical review can group draft stories by setting, character role, central conflict and final resolution. If too many sit in one cluster, the editor can reject the weak duplicates, request a genuinely different direction or commission a human-developed premise. This is an editorial judgement about a collection, rather than a mathematical formula for creativity.
The issue extends beyond novels. Game writers, advertisers and short-video makers increasingly use generators to offer quick alternatives. LiveAIWire has examined a related creative possibility in creating video games with words, where AI tools make it easier to generate interactive entertainment. Lowering the cost of producing an idea is useful only if the ideas are sufficiently varied to hold an audience’s attention.
Where machine creativity still needs a human brief
There is a productive way to respond to repetitive AI fiction. The human can decide what kind of surprise the story should contain before the generator begins to elaborate it. An editor might start with an observed conflict between neighbours, a particular place or a character whose motivations cannot be reduced to a stock description. The model can then be used to test different conversations, settings or endings against that independent starting point.
This is slower than accepting the first elaborate answer, but it creates a useful distinction between generating words and directing a story. Editors should also recognise that a good prompt cannot guarantee genuinely distinctive writing. It may take several rounds of comparison, rejection and revision to achieve a coherent voice. When the output is intended for publication, the person taking responsibility for it still needs to check originality, tone and suitability for the audience.
The lighthouse study offers a useful test rather than a reason to abandon creative tools. If a collection of supposedly unrelated stories begins to share names, images and narrative turns, it is time for the editor to ask what the machine has been choosing on autopilot.
What the lighthouse keeper reveals about AI
The image of Elias at the lighthouse sticks in the memory because it is so particular. Behind it sits a more general insight: technology that produces enormous amounts of language can still carry an unusually visible imprint of the examples that shaped its preferences.
That is not proof that humans are always more imaginative. Professional writers also have repeated habits and favourite settings. The difference is that a person can reflect on those habits, seek out new experiences and make a deliberate artistic choice. A user should demand the equivalent range of possibilities from an AI tool, rather than accept the first polished paragraph as evidence of inspiration.
If a chatbot keeps returning to the same lighthouse, the practical question is not how to ban lighthouses. It is how to make the next story truly belong to the person asking for it.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.