Research / AI Creativity, Authorship, Translation, and Co-Creation
AI Creativity, Authorship, Translation, and Co-Creation
Elkins’ work on AI creativity asks what language models can generate, imitate, translate, and co-create, and what their outputs reveal about authorship, intention, literary value, and creative practice. The work focuses on AI as a creative and aesthetic force: machine-written fiction, literary imitation, authorship, translation, human-AI improvisation, and the evaluation of generated language.
The through-line is practical and theoretical at once. Early work tested whether GPT-3 could perform recognizable creative-writing tasks. Later work asks what AI does to the author as a cultural and aesthetic category, why fiction remains difficult for transformer architectures, how literary translation can be evaluated with AI, and what live human-AI performance reveals about co-creation and improvisation.
Selected work and projects
Can GPT-3 Pass a Writer’s Turing Test?
Elkins and Chun, Journal of Cultural Analytics 5.2, 2020
Published four months after the release of GPT-3, this article proposed the writer’s Turing test as a framework for evaluating large language model creative competence. Rather than asking whether GPT-3 was intelligent in a general sense, the article tested whether it could produce recognizable literary and philosophical writing in response to humanities-relevant prompts.
The article became an early reference point for evaluating LLM creativity. It was cited by Luciano Floridi and Massimo Chiriatti in Minds and Machines for the empirical claim that “GPT-3 writes better than many people,” and by Martin Paul Eve in The Digital Humanities and Literary Studies (Oxford University Press, 2022). Later uptake includes work on machine psychology, AI ghostwriting, GPT-3 creativity testing, machine-text evaluation, plagiarism and academic writing, and Turing-test research on naturalistic communication and deception.
AI Comes for the Author
Poetics Today 45.2, 2024
This article examines what large language models do to authorship as a creative, cognitive, and aesthetic category. Rather than treating AI authorship only as a question of copyright or content generation, it asks how machine-generated language changes older assumptions about intention, originality, literary value, authority, and interpretation.
If models can generate plausible literary language, the problem is not only whether they can produce text, but what happens to the cultural and aesthetic role of authorship when production is distributed among prompts, training data, model architecture, and human evaluation.
James Phelan engages the article in Poetics Today 45.2 as part of an “instructive debate” about AI, narrative, and authorship.
The AI Fiction Paradox
Elkins, arXiv 2603.13545, 2026
This paper names the AI Fiction Paradox: contemporary AI systems are trained on massive amounts of fiction and need even more of it, yet still struggle to generate compelling fiction. The paper identifies three challenges for current architectures: narrative causation, informational revaluation, and multi-scale emotional architecture.
The argument reframes AI fiction as a problem of narrative intelligence rather than surface fluency. Fiction requires events that feel surprising as they unfold and retrospectively inevitable after they occur; details whose importance can be reweighted after the fact; and emotional patterning across words, sentences, scenes, and full narrative arcs. The paper also raises a safety question: if models eventually master fiction’s cognitive and emotional patterns, those same capacities could become powerful tools for persuasion and manipulation.
In Search of a Translator: Using AI to Evaluate What’s Lost in Translation
Elkins, Frontiers in Computer Science 6, 2024
This article develops a computational method for studying literary translation through Proust’s Du côté de chez Swann and its English translations. It uses stylometry, lexical diversity, readability, and emotional-arc analysis to examine how translation changes the experience of a literary work.
The study finds that Lydia Davis comes closest to preserving several features of Proust’s style among the human translators examined, while Mistral performs best among the LLMs evaluated at capturing emotional nuance in the French original. The article shows how AI can surface undertheorized features of translation without replacing literary judgment, especially where translation changes rhythm, lexical density, emotional arc, and the reader’s experience of style.
The article has been engaged by Huang and Cheung in Humanities and Social Sciences Communications as part of later work on AI, translation, and literary evaluation.
DIVAbot: Early Live Human-AI Improvisation with GPT
Arts at Denison, January 25, 2021
DIVAbot was a live theatrical improvisation between a human performer and a transformer-based AI, staged on the 100th anniversary of Karel Čapek’s R.U.R., the 1921 play that coined the word “robot.” Created with Jon Chun, Jim Dennen, and Lauren Katz, the performance used GPT-2-based language generation to test whether a machine could participate in live theatrical play.
The project took place twenty-two months before ChatGPT and before later European AI-theatre productions such as THEaiTRE’s Prague premiere and the Young Vic’s AI. It treats creativity as interaction: not simply whether a model can produce text, but whether human performers and audiences can make meaning with machine-generated language in real time.
Public Engagement on AI and Creativity
Interviews, panels, talks, and media appearances
Elkins’ public-facing work on AI creativity includes interviews, panels, talks, and media appearances on generative AI, the arts, machine authorship, translation, and creative labor. These appearances connect technical changes in language generation to broader public questions about artistic practice, authorship, education, and the future of creative work.