Beyond the Stochastic Parrot: A Test for Narrative Creativity in Large Language Models

Debates about creativity in large language models tend to focus on mechanism (whether the model "really" understands, intends, or invents) and consequently stall. This paper proposes shifting the question from mechanism to function: does the output work as a story? I introduce a diagnostic task grounded in improvisational theatre practice and narrative theory. A canonical scene with a strongly compressed cultural script (a marriage proposal) is broken at exactly one point (the box contains a shark tooth), and the model is asked to continue. Responses are classified along three levels of narrative integration: the break is ignored (L1), acknowledged but inert (L2), or made structurally necessary (L3). A substitution test distinguishes L2 from L3. A second dimension records whether integration proceeds through backstory construction (reverse causality, what improvisers call justification) or future projection (forward causality). An illustrative run across seven models of varying scale shows that no model ignored the break, that most converged on a shared secondary cliché (ocean and diving imagery), and that only the largest model produced a non-substitutable object. The contribution is the method, not the ranking: the results reported here are single samples and are presented as a demonstration, not as evidence. I outline what a properly powered study would require and argue that the framework generalises to any sufficiently canonical scene. Revised and extended from an essay first published at https://thomas-brandau.com/beyond-the-stochastic-parrot-a-test-for-narrative-creativity-in-large-language-models/ on 30 May 2026.