From Pixels to Personas: Investigating and Modeling Self-Anthropomorphism in Human-Robot Dialogues

TL;DR

The paper investigates self-anthropomorphism (SA) in human–AI dialogues, distinguishing SA from non-self-anthropomorphic (NSA) responses and examining ethical implications across embodiments. It introduces Pix2Persona, a dataset with 143K dialogue turns that pairs original bot responses with SA and NSA variants, enabling systematic study and dynamic SA control. A GPT-4–based classifier validates SA detection (P=81.68%, R=82.57%, F1=81.92%; κ=0.83), and analyses reveal task-driven SA prevalence and distinct linguistic markers. The authors propose a Dual-Capability model (fine-tuned Mistral-7B) and demonstrate transformations SA↔NSA across multiple tasks, achieving competitive performance and enabling adaptive SA levels tailored to ethical and user-experience considerations. Overall, the work provides tools and insights for ethically managing SA in embodied AI, with implications for safer, more engaging human–robot interactions and future research on context-aware anthropomorphism.

Abstract

Self-anthropomorphism in robots manifests itself through their display of human-like characteristics in dialogue, such as expressing preferences and emotions. Our study systematically analyzes self-anthropomorphic expression within various dialogue datasets, outlining the contrasts between self-anthropomorphic and non-self-anthropomorphic responses in dialogue systems. We show significant differences in these two types of responses and propose transitioning from one type to the other. We also introduce Pix2Persona, a novel dataset aimed at developing ethical and engaging AI systems in various embodiments. This dataset preserves the original dialogues from existing corpora and enhances them with paired responses: self-anthropomorphic and non-self-anthropomorphic for each original bot response. Our work not only uncovers a new category of bot responses that were previously under-explored but also lays the groundwork for future studies about dynamically adjusting self-anthropomorphism levels in AI systems to align with ethical standards and user expectations.

Figures (15)

• Figure 1: Samples in the Pix2Persona dataset. (a) Transform self-anthropomorphic responses to non-self-anthropomorphic. (b) Transform non-self-anthropomorphic responses to self-anthropomorphic.
• Figure 2: Classifier prompt defining self-anthropomorphic and non-self-anthropomorphic bot responses. Definitions are in blue, and placeholders within orange-highlighted text correspond to a single dialogue turn. See Appendix \ref{['classify_example']} for detailed explanation.
• Figure 3: Trends in self-anthropomorphic bot responses across various datasets. The bar chart shows the ratio of dialogue turns classified as self-anthropomorphic out of 100 sampled turns from each dataset, with each bar color indicating a different dialogue task type.
• Figure 4: Point biserial correlation coefficients between word categories and self-anthropomorphism labeling in bot responses. A positive value indicates a tendency for the word category to be associated with self-anthropomorphic responses (labeled as 1), while a negative value suggests an association with non-self-anthropomorphic responses (labeled as 0).
• Figure 5: The number of responses categorized as SA in original SA responses, naive bot responses, and transformed NSA responses across five open-domain dialogue datasets.