LLMs and TTRPGs

Study title: Evaluating the effects of language model player assistants on the social dynamics of table-top role-playing game parties.

This was a summer project that I conducted as part of the Foundation Year of my Interactive AI CDT PhD. It involved the development of a LLM-based chatbot that would aid player during Dungeons and Dragons gameplay. The motivation for this project was a recognition of the need to streamline information retrieval from the various online resources available for TTRPG players. The project was aimed at facilitating greater socialisation between players and increasing immersion in the campaign, by reducing the amount of time spent on retrieving factual information regarding the game rules, characters and lore. To this end, play sessions were carried out of a D&D one-shot campaign with three players and one Dungeon Master. Contrary to initial conjecture, the primary uses of the chatbot were roleplay inspiration, which supported the theatrical choices that players made about their character. Despite the hardware limitations present, it was the players’ ability to generate effective prompts which had a greater impact on the generated answers. Particularly, the issue of excessive safeguarding arose, in which the prompt from one of the players about a combat action generated a warning response from the chatbot about doing harm to animals (for context the prompt was asking whether it was possible to smash a rat using another rat). These findings leav space for future work in developing assistive tools that can more effectively infer user intent and generate useful responses or provide active choices for players as the state of the game changes.