Abstract
Language models are increasingly attracting interest from writers. However,such models lack long-range semantic coherence, limiting their usefulness forlongform creative writing. We address this limitation by applying languagemodels hierarchically, in a system we call Dramatron. By building structuralcontext via prompt chaining, Dramatron can generate coherent scripts andscreenplays complete with title, characters, story beats, locationdescriptions, and dialogue. We illustrate Dramatron's usefulness as aninteractive co-creative system with a user study of 15 theatre and filmindustry professionals. Participants co-wrote theatre scripts and screenplayswith Dramatron and engaged in open-ended interviews. We report criticalreflections both from our interviewees and from independent reviewers whowatched stagings of the works to illustrate how both Dramatron and hierarchicaltext generation could be useful for human-machine co-creativity. Finally, wediscuss the suitability of Dramatron for co-creativity, ethical considerations-- including plagiarism and bias -- and participatory models for the design anddeployment of such tools.