Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cooking Up Politeness in Human-AI Information Seeking Dialogue

David Elsweiler, Christine Elsweiler, Anna Ziegner

TL;DR

The paper investigates how user politeness influences GenAI information-seeking in a task-oriented cooking domain. It combines empirical analysis of 30 Wizard-of-Oz cooking dialogues to identify politeness-based user clusters and large-scale LLM–LLM simulations across five politeness profiles and three open-weight models. Findings show that politeness systematically affects response length, information nuggets, and energy efficiency, with engaging styles increasing output and costs, while impolite style reduces efficiency; diminishing returns emerge as verbosity grows. The work highlights fairness and sustainability considerations for designing inclusive conversational agents and provides a foundation for policy-oriented guidelines and future cross-domain validation.

Abstract

Politeness is a core dimension of human communication, yet its role in human-AI information seeking remains underexplored. We investigate how user politeness behaviour shapes conversational outcomes in a cooking-assistance setting. First, we annotated 30 dialogues, identifying four distinct user clusters ranging from Hyperpolite to Hyperefficient. We then scaled up to 18,000 simulated conversations across five politeness profiles (including impolite) and three open-weight models. Results show that politeness is not only cosmetic: it systematically affects response length, informational gain, and efficiency. Engagement-seeking prompts produced up to 90% longer replies and 38% more information nuggets than hyper-efficient prompts, but at markedly lower density. Impolite inputs yielded verbose but less efficient answers, with up to 48% fewer nuggets per watt-hour compared to polite input. These findings highlight politeness as both a fairness and sustainability issue: conversational styles can advantage or disadvantage users, and "polite" requests may carry hidden energy costs. We discuss implications for inclusive and resource-aware design of information agents.

Cooking Up Politeness in Human-AI Information Seeking Dialogue

TL;DR

The paper investigates how user politeness influences GenAI information-seeking in a task-oriented cooking domain. It combines empirical analysis of 30 Wizard-of-Oz cooking dialogues to identify politeness-based user clusters and large-scale LLM–LLM simulations across five politeness profiles and three open-weight models. Findings show that politeness systematically affects response length, information nuggets, and energy efficiency, with engaging styles increasing output and costs, while impolite style reduces efficiency; diminishing returns emerge as verbosity grows. The work highlights fairness and sustainability considerations for designing inclusive conversational agents and provides a foundation for policy-oriented guidelines and future cross-domain validation.

Abstract

Politeness is a core dimension of human communication, yet its role in human-AI information seeking remains underexplored. We investigate how user politeness behaviour shapes conversational outcomes in a cooking-assistance setting. First, we annotated 30 dialogues, identifying four distinct user clusters ranging from Hyperpolite to Hyperefficient. We then scaled up to 18,000 simulated conversations across five politeness profiles (including impolite) and three open-weight models. Results show that politeness is not only cosmetic: it systematically affects response length, informational gain, and efficiency. Engagement-seeking prompts produced up to 90% longer replies and 38% more information nuggets than hyper-efficient prompts, but at markedly lower density. Impolite inputs yielded verbose but less efficient answers, with up to 48% fewer nuggets per watt-hour compared to polite input. These findings highlight politeness as both a fairness and sustainability issue: conversational styles can advantage or disadvantage users, and "polite" requests may carry hidden energy costs. We discuss implications for inclusive and resource-aware design of information agents.
Paper Structure (37 sections, 2 figures, 2 tables)

This paper contains 37 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Variation in linguistic politeness
  4. Language in Information Interaction
  5. Analysing Frummet et al's data
  6. Annotation and Reliability
  7. Validating the Annotation Process
  8. User Representation and Clustering
  9. Visualisation and Interpretation
  10. Simulating (Im)polite Conversations
  11. Simulation Approach
  12. Recipe task domain.
  13. User profiles.
  14. Agent behaviour.
  15. Conversation flow.
  16. ...and 22 more sections

Figures (2)

  • Figure 1: Z-score standardised code frequency across clusters
  • Figure 2: H3a/H3b efficiency diagnostics. Top: nuggets vs. words; bottom: nuggets vs. energy (Wh). Columns are agents; colours denote user clusters. LOESS smooths with 95% CIs over step-level points. X-axes are free per column to reflect very different scales. Patterns show early gains followed by plateaus (diminishing returns).