Presenting Large Language Models as Companions Affects What Mental Capacities People Attribute to Them

Paper Title

Presenting Large Language Models as Companions Affects What Mental Capacities People Attribute to Them

Publication Info

• Topic area: Human-AI interaction and public perception of AI systems.
• Keywords: Large language models, mental capacities, anthropomorphism, AI messaging, public perception, experimental studies, reliance behaviors, information retrieval, cognitive capacities, emotional capacities.

Background and Problem

• Problem / challenge: There is limited understanding of how public messaging about AI systems, specifically large language models (LLMs), influences beliefs about their mental capacities and usage behaviors.
• Significance: Understanding these effects is critical as LLMs become more integrated into society, shaping how people interact with and rely on these systems in various domains.
• Motivation and related work: Prior research has shown that anthropomorphic design and messaging can affect perceptions of AI systems, but most studies focus on hypothetical or custom-made systems. This paper addresses the gap by exploring the impact of real-world messaging about broadly accessible LLMs.

Solution

• Proposed approach: Experimental studies investigating the effects of presenting LLMs as machines, tools, or companions on people's beliefs and reliance behaviors.
• Novelty:
  1. Demonstrates that presenting LLMs as companions increases attributions of cognitive and emotional capacities.
  2. Shows that presenting LLMs as machines reduces reliance on inconsistent LLM responses.
  3. Provides evidence that messaging effects persist across time and are moderated by task familiarity.
  4. Explores nuanced impacts of messaging on reliance behaviors in information retrieval tasks.
• Procedure and key techniques:
  • Study 1: Participants (N = 470) watched videos presenting LLMs as machines, tools, companions, or no video, followed by a survey assessing mental capacity attributions and additional beliefs.
  • Study 2: Nine months later, participants (N = 604) completed factual question-answering tasks using pre-generated LLM responses, followed by a survey measuring reliance behaviors and mental capacity attributions.
  • Statistical analyses included linear mixed-effects regression and exploratory factor analysis.

Results

• Concrete findings:
  • Participants exposed to the "companion" video attributed higher cognitive (5.00 vs. 4.47) and emotional (2.15 vs. 1.62) capacities to LLMs compared to other groups.
  • Participants who watched the "machine" video relied less on inconsistent LLM responses (57.05% agreement vs. 62.33% overall).
  • The effect of messaging on mental capacity attributions persisted across time but was moderated by task familiarity, reducing cognitive attributions after completing factual tasks.
• Advantage over baselines:
  • Companion messaging increased mental capacity attributions beyond baseline beliefs.
  • Machine messaging fostered skepticism in detecting unreliable outputs.
• Experiments / evaluation:
  • Study 1: Survey measured attribution of 40 mental capacities and five additional beliefs (e.g., trust, human-likeness).
  • Study 2: Factual QA tasks tested reliance on LLM responses with varying correctness and consistency.
  • Metrics included agreement with LLM answers, confidence, follow-up questions, and time spent.
• Limitations and future work:
  • Limited ecological validity due to controlled messaging and task design.
  • Narrow scope of reliance behaviors studied; future work should explore long-term interactions and diverse tasks.
  • Need for longitudinal studies to understand cumulative effects of repeated or conflicting messages.

Summary

This paper investigates how presenting large language models (LLMs) as machines, tools, or companions influences public beliefs about their mental capacities and reliance behaviors. Experimental results show that companion messaging increases attributions of cognitive and emotional capacities, while machine messaging reduces reliance on inconsistent LLM responses. These effects were replicated across time and moderated by task familiarity. The findings highlight the importance of public messaging in shaping perceptions and interactions with AI systems, suggesting that anthropomorphic presentations may increase expectations, while mechanistic presentations may foster skepticism. Future research should explore real-world messaging dynamics and broader reliance behaviors in varied contexts.