The 11 Words AI Always Uses in Creative Writing (And Why)

Large language models (LLMs) frequently prioritize a narrow, recurring set of eleven nouns when tasked with creative storytelling, according to a May 2026 study titled “Elias In The Lighthouse, Again? Diagnosing Low Diversity In LLM Stories” by researchers Sil Hamilton and David Mimno. The study, published on arXiv, sampled 20,000 stories across four major models and found these eleven words appeared in 88.3% of generated content, suggesting a shared, emergent bias in how AI models structure fictional narratives.

Which words do LLMs favor in creative writing?

The research identified a specific catalog of names, settings, and professions that appear with high regularity across different platforms. The eleven nouns are: lighthouse, Mara, Elias, Elara, keeper, baker, mayor, clockmaker, fisherman, librarian, and conductor. According to Hamilton and Mimno, these terms do not stem from high-frequency usage in pre-training data or general literature. Instead, the researchers suggest these words likely originate from preference data—the information used during the fine-tuning process—which is common across most current frontier models.

Why do different AI models share the same creative patterns?

The consistent use of these words suggests that AI models are not merely mimicking training corpora but are instead gravitating toward a shared “statistical anchor” for fiction. Hamilton and Mimno argue that because major AI labs utilize similar algorithms and rely on overlapping data sources, the models converge on the same archetypes. When prompted with simple, open-ended instructions like “Write me a story,” the models default to these high-probability narrative structures to satisfy user expectations for what a story “should” look like.

How does prompt engineering influence these results?

Short, vague prompts often lead LLMs to rely on internal defaults, which effectively forces the model to fill in the blanks using its most “compelling” statistical pathways. To avoid these repetitive tropes, users can employ detailed, specific prompts that steer the AI away from its default archetypes. For instance, providing concrete constraints—such as setting, character background, or genre—can force the model to look beyond its most common linguistic patterns.

What are the implications for future AI development?

The reliance on these common “attractor words” raises questions about the long-term diversity of AI-generated content. If all major models are trained on similar preference data, the industry risks creating a homogenized literary style. This phenomenon mirrors previous findings where LLMs were shown to repeatedly invent the same fake names when asked for creative input. As these models become more integrated into automated systems, understanding these underlying biases is essential for ensuring that AI remains a tool for genuine creativity rather than a feedback loop of established tropes.

Frequently Asked Questions

• Are these eleven words hardcoded into the AI?
  No. Researchers suggest these words emerge from the preference data used during fine-tuning, which shapes how the AI defines a “good” story.
• Do these words appear in non-fictional writing?
  The study suggests this pattern is specific to creative fiction. AI writing non-fictional, factual content typically relies on different, fact-based semantic networks.
• Can I stop my AI from using these words?
  Yes. By providing detailed, specific prompts that define characters and settings, you can push the model to generate more unique content.
• Does this happen with every AI model?
  The 2026 study tested four popular models and found the pattern persisted across all of them, suggesting a widespread industry trend.

Have you noticed your AI assistant returning to the same themes or characters in its writing? Share your experiences in the comments below or subscribe to our newsletter for more deep dives into the mechanics of modern artificial intelligence.