Dienstag, 14. Oktober 2025, 19:00 Uhr
Geiwi-Turm, SR 4O432
Abstract: Over the past few years, large language models (LLMs) like ChatGPT have become widely used, with generative AI (GenAI) systems now a part of many people's daily lives. Recently, user politeness in interactions with GenAI has received public attention, with claims that politeness causes increased energy consumption leading experts to discourage people from being polite in their interactions with GenAI. Yet, to date we know little about user politeness patterns in human-GenAI interactions and to what degree these vary. As variation in user politeness may not only affect energy consumption but also the length and quality of system replies, we address these aspects in our study.
In this talk, we are going to present the findings of our interdisciplinary study combining linguistic and information science approaches. It explores variation in user politeness in task-oriented information-seeking interactions with generative AI systems to then establish whether different politeness profiles influence response length, information transfer and energy efficiency. In the linguistic part of the study, we annotated Frummet et al.’s (2024) publicly available Cooking with Conversation dataset for politeness-relevant speech acts, adapting and expanding Leech’s (2014) classification scheme. Based on the annotated conversations, we conducted a cluster analysis, from which we derived four distinct politeness clusters reflecting a continuum of interactional styles ranging from highly respectful, appreciative and engaged to task-oriented and efficient. In the information science part of the study, we simulated ca. 18,000 user–agent cooking conversations to examine how different politeness profiles influence system behaviour. Our analysis revealed that user politeness significantly impacts response length, the amount of information conveyed, and energy efficiency, with more engaged and polite styles eliciting longer, more informative replies but mostly at a higher energy cost. These findings highlight how user interaction styles shape both the content and resource demands of generative AI systems, offering insights for designing more efficient and effective interactions.
In the concluding part of our talk, we will outline plans for a larger interdisciplinary project that will investigate both inter- and intralingual variation in user politeness during different taskoriented human-GenAI interactions, examining its impact on factors such as user satisfaction, system inclusivity, and information transfer.
Alle Interessierten sind herzlich eingeladen!