Repetition and Defamiliarization in AI Dungeon and Project December

Abstract (in English): 

Recent advances in machine learning provide new opportunities for the exploration of creative, interactive works based around generative text. This paper compares two such works, AI Dungeon (Walton 2019) and Project December (Rohrer 2020), both of which are built on the same artificial intelligence (AI) platform, OpenAI’s GPT-2 and GPT-3. In AI Dungeon, the player can choose from several predetermined worlds, each of which provide a starting point for the story generation. However, while interacting with the system within this world, the player can stop, edit, modify and retry each utterance, allowing the player to “sculpt” the AI’s responses, and choose what goes into the AI’s memory, helping to shape the overall direction of the story. At a broader level, the player can edit world descriptions, insert scripts between the AI and the player (themselves or others), and share these worlds/scenarios with other players. Similarly, in Project December, the player interacts with several AI “matrices”, either directly through conversations, or more indirectly by creating new matrices by defining a starting paragraph and sample responses, which can then be “spun up”, tested, and tweaked much like the worlds in AI Dungeon. These matrices can also be shared with other players.

When interacting with both works, there is a need for the player to repeatedly engage with the work to learn how to entice a satisfying experience from the system (Mitchell 2012; 2020). However, the key difference is the framing of the experience. In AI Dungeon the person experiencing the work is either taking on the role of the player, entering text and seeing how the AI responds, or that of an author or perhaps a co-author, tweaking the input to the AI or its responses or adjusting the underlying scenario to get a desired response. In contrast, Project December is presented as part of a fictional website for a “Project December” run by “Rhinehold Data Systems”, promising the opportunity to talk to “the world’s most super computer”. Upon accessing the “customer terminal”, which looks and feels like an old dialup terminal, the player takes on the loosely defined role of “Professor Pedersen” whose “.plan file”, dated November 13, 1982, contains several tasks related to the various “matrices”, suggesting a mystery to be solved and a larger narrative to be explored. I will argue that whereas AI Dungeon attempts to provide players with access to and an uncritical understanding of how the underlying AI system works, Project December’s narrative framing instead defamiliarizes the play experience (Mitchell et al. 2020), potentially creating a more emotional connection between the player and the “matrices”, and thereby encouraging the player to critically reflect on the implications of the underlying technological platform.


ELO 2021: Artificial and generated text, 27 May 2021

The permanent URL of this page: 
Record posted by: 
Daniel Johannes...