Collective Memory Lapse: LLMs Omissions in Roleplay

This paper presents an artistic and Digital Humanities case study examining how Large Language Models represent historical trauma when used as game masters in role-playing scenarios. The authors designed an experimental, text-based time-travel game about the 1945 massacre of Hungarian Jews at Rechnitz, Austria, and tested it alongside an earlier prototype on the 1936 murder of philosopher Moritz Schlick. Players could select among European, Chinese, and US-American LLMs, and the games were deliberately designed to be non-immersive, with the models instructed to adhere strictly to historical facts and to offer four numbered choices after each scene, prompting frequent collective debriefing.

The methodology combined a three-month public installation in Vienna in 2025, qualitative interviews and debriefings with players, and a quantitative analysis of 115 LLM-generated role-play texts using natural language processing techniques such as semantic similarity and sentiment analysis. Players were grouped into three distinct user types, and the text analysis revealed significant differences in how the individual models presented the same historical content. The authors also conducted a self-experiment in which they modified the prompt to prevent the player from altering history, finding that the model consistently avoided depicting the violence directly and offered only vague, euphemistic language, citing safety and content rules when questioned.

The authors conclude that these omissions trivialized the atrocity and echoed the postwar Austrian pattern of rumor and silence that public discourse had already worked to overcome, suggesting that generative AI can reintroduce distortions into collective memory. The work raises open questions about which historical events models know and how they present them, what guidelines govern their filters, and what it means for societal memory when such editorial decisions rest with AI companies.