Narrative and Text Generation AI

Submitter: Addison Eldin, U of Pittsburgh

——————————————————

The experiment:

Students and I discussed AI in a course called Narrative and Technology, where the focus was on how narrative studies scholarship has engaged with narrative across a variety of media. In the class prior to the main exercise described here, we discussed video game AI in the context of 2005’s Façade, which uses non-generative AI as part of its text input parser and “drama manager,” where the freeform user input is interpreted using language parser rules and then used to choose and sequence relevant narrative beats to carry the story toward its many possible endings.

Following this class, we looked at AI Dungeon, a roleplaying game that users interact with much like ChatGPT through an input and generated output back-and-forth. The game itself relies on ChatGPT to act as a kind of storyteller or game/dungeon master to build a story scenario that users can interact with by typing out statements by their character, or describing actions their character wishes to take. After a discussion about the efficacy of the technology in maintaining narrative structure and accurately interpreting/building on the open-ended user input, I then provided a brief lecture on the fundamentals of generative pre-trained transformer models of text generation AI. This lecture then enabled a deeper conversation about whether the narrative text produced between AI Dungeon and the user could be considered a narrative, or how the narrative it produces is different from other forms of interactive narrative.

Results:

Students, prior to the discussion about generative pre-trained transformer technology, had mixed reactions to the story produced by their interactions with AI Dungeon. Some noted that it was surprisingly adept at responding to input and loosely keeping track of the relevant narrative elements. Many more noted the limitations the technology placed on narrative exploration, and that the product lacked what might be called a sense of “narrativity.” Some students drew comparison to tabletop roleplaying systems (e.g. Dungeons and Dragons). After we discussed generative pre-trained transformer technology by breaking down each term, with a focus on the stochastic nature of something like ChatGPT, students seemed split on whether the result of their interaction with such technology could rightly be considered narrative.

We also discussed how fundamentally different something like AI Dungeon is from something like Facade, with its pre-constructed set of story beats and responses to user input, and how that informs what role interactivity has in the study of narrative. Finally, we discussed whether students found the future narrative possibilities of something like ChatGPT-as-game master were exciting or alienating/dehumanizing. Each of these conversations demonstrated that students were not only interested in AI Dungeon as a narrative system despite its flaws, but they were further interested in the consequences of such technology on the future of narrative and human interaction.

Relevant resources: https://addeldin.github.io/projects/narrative-and-text-generation-ai/

Contact: addison.eldin[AT]pitt[DOT]edu