A Taxonomy of Linguistic Expressions That Contribute To Anthropomorphism of Language Technologies

Research Background and Issues

• Problem Identification: In the realm of fintech and large language models (LLMs), people often mistakenly project certain human traits onto non-human entities and objects. This phenomenon of "anthropomorphism" can introduce numerous potential risks and negative impacts on how users interact with and utilize these technologies. Specifically, the authors highlight how the textual outputs of language technologies can mislead users into perceiving them as human-like entities, necessitating a deeper understanding of the causes and mechanisms behind this influence.
• Significance of the Problem: As LLMs (e.g., ChatGPT, Claude) become increasingly prevalent, this anthropomorphism not only risks fostering excessive reliance, emotional attachment, or overestimations of technological capabilities but also raises profound implications for social relationships, privacy, and the attribution of technological responsibility.
• Research Motivation and Related Work: The authors note that while numerous existing studies have explored intentional or unintentional anthropomorphism in technology design and its resulting "misleading trust," there is a lack of a systematic framework to describe the diverse forms of anthropomorphic expressions in language outputs and their potential impacts. This study aims to advance understanding in this area to support more responsible design of language technologies.

Solution

• Methods and Framework:
  1. The authors propose a "taxonomy" of anthropomorphic language expressions, focusing on 19 types of expressions in textual outputs that can evoke anthropomorphic perceptions, alongside 5 "guiding lenses" to identify the mechanisms behind these expressions.
  2. Data Source: A collection of 395 real-world cases from existing literature and public examples provides empirical support for the proposed taxonomy.
• Innovations:
  1. The study introduces the first detailed classification system for anthropomorphic mechanisms in language technology outputs, describing key traits such as "suggestions of internal states, social positioning, materiality, autonomy, and communication skills."
  2. It systematically integrates related categories from existing literature while expanding on newly identified forms of expression through case analysis.
• Implementation Steps:
  1. The authors designed a participatory and open-coding content analysis method to classify textual outputs from real-world interaction cases.
  2. They established several "guiding lenses," such as suggestive of internal states, social positioning, etc., to frame which output expressions might induce anthropomorphism, further refining these into 19 specific expression types.
  3. Data and method validation: Multiple iterative analyses of cases were conducted until the taxonomy reached content saturation.

Research Findings

• Specific Findings:
  1. Five Guiding Lenses: These help identify dimensions of anthropomorphism in language technology outputs, such as "suggestions of internal states" (e.g., perceived emotions or awareness), "materiality" (whether sensory or bodily experiences are implied), "autonomy," and "communication skills" (e.g., sentence variations, dialogue patterns).
  2. Nineteen Expression Types: These encompass categories such as expressions of intelligence, emotional expression, self-awareness, intention, self-evaluation, and social relationships. For example, a machine-generated dialogue like "I want to learn more about the world" might involve dual anthropomorphic cues of "emotional suggestion" and "autonomous behavior."
• Comparative Advantages:
  1. The taxonomy's broad coverage and in-depth analysis of real-world data enable researchers and developers to precisely identify potential "misleading anthropomorphic" expressions in textual outputs.
  2. The taxonomy can serve as a benchmark for measuring the degree of humanization in language outputs, supporting more precise scientific experiments and user studies.
• Experimental or Evaluation Results:
  1. Case analyses validated the impact of different types of language expressions on users' anthropomorphic perceptions. For instance, does explicitly stating a robot's identity reduce anthropomorphic perceptions? (The study found that "expressing limitations and self-descriptions of identity often simultaneously reinforce human-like perceptions.")
  2. Data revealed that anthropomorphism is pervasive across various technological scenarios/platforms, but its degree is context-dependent, particularly prominent in perceptions of emotional communication, ethical judgment, and behavioral autonomy.
• Limitations and Future Directions:
  1. The study's cases are primarily based on English contexts, limiting the taxonomy's cross-linguistic and cultural generalizability (e.g., how multilingual technologies evoke anthropomorphism in different contexts requires further exploration).
  2. The study does not delve into how models might generate more cautious, non-misleading outputs based on the taxonomy.
  3. Future research could investigate user behavioral responses to "anthropomorphism" across different backgrounds and explore how anthropomorphic language might be restricted in appropriate contexts to avoid potential societal biases.

Conclusion

This study significantly advances both practical and theoretical understanding of the "perceptual impacts of anthropomorphic language" by developing a large-scale analysis framework for LLM outputs based on real-world interactions. The taxonomy provides a practical guide for future generative AI design and human-computer interaction research while highlighting the critical connections between anthropomorphism, ethics, and technological boundaries, with far-reaching implications.