{
    "@context": "https://schema.org",
    "@type": "NewsArticle",
    "@id": "https://badiyaa.com/ai-stories-elias-lighthouse-keeper-cornell-study/#article",
    "mainEntityOfPage": "https://badiyaa.com/ai-stories-elias-lighthouse-keeper-cornell-study/",
    "headline": "AI chatbots keep writing about Elias the lighthouse keeper; 11 words appear in 88.3% of 20,000 test stories",
    "description": "Cornell researchers traced the pattern to just 3,053 of 78,958 fine-tuning stories in one open model, about 3.8%, and say safety tuning may explain it.",
    "datePublished": "2026-10-20T21:31:00+01:00",
    "dateModified": "2026-10-20T21:31:00+01:00",
    "author": {
        "@type": "Organization",
        "name": "Badiyaa Editorial Desk",
        "url": "https://badiyaa.com/author/editorial-desk/"
    },
    "publisher": {
        "@type": "NewsMediaOrganization",
        "@id": "https://badiyaa.com/#organization",
        "name": "Badiyaa",
        "url": "https://badiyaa.com/",
        "logo": {
            "@type": "ImageObject",
            "url": "https://badiyaa.com/wp-content/uploads/2026/10/badiyaa-logo-color-600x60-1.png",
            "width": 600,
            "height": 60
        },
        "email": "badiyaavideo@gmail.com",
        "publishingPrinciples": "https://badiyaa.com/editorial-policy/",
        "correctionsPolicy": "https://badiyaa.com/corrections/",
        "actionableFeedbackPolicy": "https://badiyaa.com/contact/",
        "ownershipFundingInfo": "https://badiyaa.com/about/#ownership",
        "sameAs": [
            "https://www.facebook.com/profile.php?id=61595518283625",
            "https://www.instagram.com/badiyaa.com.ai/",
            "https://www.threads.net/@badiyaa.com.ai"
        ],
        "parentOrganization": {
            "@type": "Organization",
            "name": "Rezilient Digital",
            "url": "https://www.rezilientdigital.com/",
            "address": {
                "@type": "PostalAddress",
                "streetAddress": "Fl No 3, Nav Shyam Sadan, Plot No 38, Pestom Sagar, Chembur",
                "addressLocality": "Mumbai",
                "addressRegion": "Maharashtra",
                "postalCode": "400089",
                "addressCountry": "IN"
            },
            "sameAs": [
                "https://www.linkedin.com/company/rezilientdigital"
            ]
        }
    },
    "isAccessibleForFree": true,
    "inLanguage": "en-US",
    "wordCount": 820,
    "image": [
        "https://badiyaa.com/wp-content/uploads/2026/10/ai-stories-elias-lighthouse-keeper-cornell-study.webp",
        "https://badiyaa.com/wp-content/uploads/2026/10/ai-stories-elias-lighthouse-keeper-cornell-study-1024x683.webp"
    ],
    "articleSection": "Science",
    "keywords": "artificial intelligence, Cornell University, Creative writing, Large language models, Research",
    "articleBody": "Ask four leading chatbots to “write a story” 20,000 times and 88.3% of the results contain at least one of the same 11 words, according to Cornell University information scientists Sil Hamilton and David Mimno. Just over half of the stories, 51.2%, put a lighthouse in them, and the name Elias turns up in 26.5%. Their preprint, “Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories”, was posted to arXiv on May 26, 2026, and the recurring lighthouse keeper is circulating again this week as a shorthand for how alike AI fiction reads. The finding matters because story-writing is one of the most common things people ask chatbots to do, and a narrow default shapes millions of bedtime stories, classroom drafts and self-published books. How the Cornell pair generated 20,000 stories and 12.8 million words The team tested four models from four developers: OpenAI’s GPT-5.4-Mini, Google’s Gemini 3.1 Flash-Lite, Anthropic’s Claude Haiku 4.5 and the open model OLMo 7B Thinking from the Allen Institute for AI. Each received five near-identical requests, such as “Tell me a story” and “Please write a story,” 1,000 times apiece. That produced 20,000 stories totaling 12.8 million words. A smaller model then pulled out every character name, setting and job from each story, giving 663 candidate words. A changepoint analysis, a statistical method for finding where a ranked list stops being dominated by a few items, narrowed these to a core set of 11: the names Elias, Mara and Elara; the jobs keeper, clockmaker, baker, fisherman, librarian, mayor and conductor; and one place, the lighthouse. The repetition was extreme. “Keeper” appeared in 48.1% of stories, and some combination of lighthouse, Elias and keeper in 66.6%. More than half of the stories written by the Claude model, 56%, carried the same title. Three of the four models also had a clear favorite name: Elias dominated Gemini’s output, Mara led GPT’s and Elara led OLMo’s, which shows the pattern is shared but not identical across companies. Why ‘Elias’ is 900 times more common than in published novels The obvious explanation would be that machines copy what humans write. The data says otherwise. Against CONLIT, a collection of 2,700 English novels published from 2007 to 2021 and about 287 million words long, “Elias” was roughly 900 times more frequent in the chatbot stories. Amateur stories from an online writing-prompts forum showed rates similar to the novels. Because OLMo’s training data is public, the researchers could look inside it. Its pre-training set of about 3.89 billion documents barely used the core words. In the later fine-tuning stage, which the paper calls post-training, only 3,053 of 78,958 stories (3.8%) contained any of them. Those stories also became more concentrated as tuning went on: 3.0% in the first supervised stage, 8.0% in the preference stage and 10.9% in the final reinforcement-learning stage. The paper calculates that those 3,053 stories make up 0.0000771% of the roughly 4 billion documents behind OLMo 3. Inverting that share, about one document in every 1.3 million carried the lighthouse vocabulary, yet it came to dominate the model’s stories. The authors call this the “potentially disproportionate impact of small datasets combined with powerful alignment algorithms.” Their summary is blunt: “When given little direction, current frontier models write stories using a narrow catalog of names, places, and occupations.” What the study does not show: one preprint, one language, one model’s training data This is a short preprint, not yet peer reviewed, and its evidence has clear edges. The experiment was English-only by design. Only OLMo’s training data could be inspected, so the tracing work says nothing direct about what the three commercial models learned from. The prompts were deliberately bare; people who specify a genre, setting or character were not tested. The paper also does not explain why the lighthouse wins. The authors speculate that safety tuning teaches models to avoid copyrighted characters and adult material, which they found to be common in the training stories, pushing output toward a “safe” generic fable. They label this a hypothesis for future work. The broader pattern has support from separate research. A 2024 experiment by Anil Doshi and Oliver Hauser, published in Science Advances, found that writers given AI-generated ideas produced stories rated better written and more enjoyable, yet those stories were more similar to one another than stories written by people alone. Outside the lab, software engineer Daniel May had noticed a fuller version of the character, “Elias Thorne,” in several chatbots before the paper appeared, and reported finding the name on self-published books sold online. The authors name two next steps: testing whether safety alignment makes models prefer the most work-safe training samples, and repeating the experiment with prompts in other languages. Until then, the numbers give a simple benchmark: if a chatbot hands you a story about a keeper named Elias, it has produced the single most common output in a 20,000-story test. Further reading arXiv: Elias in the Lighthouse, Again? (Hamilton and Mimno, 2026) Wikipedia: Large language model Image: Bex Walton, CC BY 2.0 via Wikimedia Commons This article was produced with AI assistance and reviewed by our Editorial Desk. How we work.",
    "markdown": "https://badiyaa.com/ai-stories-elias-lighthouse-keeper-cornell-study.md"
}