This study investigates linguistic anthropomorphism in large language models (LLMs) by analyzing over one million conversational turns generated by 25 state-of-the-art models, including GPT-4, Claude, Vicuna, Llama-2, and others. The research examines anthropomorphic language markers across eight major languages and explores variation in anthropomorphic language use by model and language.
Using the Large Model Systems (LMSYS) Chat 1M dataset from Hugging Face, the study examines anthropomorphic language markers such as first-person pronouns, explicit empathy statements, and emojis across eight major languages. The methodology combines the detection of predefined anthropomorphic and empathetic phrases, the quantification of first-person references, and sentiment analysis to explore variation in anthropomorphic language use by model and language.
Results reveal substantial differences in the prevalence of anthropomorphic expressions and affective tone across LLMs, with certain models displaying significantly higher rates of empathetic and self-referential language. The analysis further identifies linguistic and cultural effects, showing that the propensity for anthropomorphic responses is not uniformly distributed across languages or cultural contexts. These findings support the theoretical concept of Situated Linguistic Anthropomorphism, according to which human-like AI language emerges from the interaction between model design and communicative setting.
The findings provide new insights into the design and evaluation of conversational AI, emphasizing the relevance of anthropomorphic cues for user trust and engagement in multilingual settings. The study contributes to understanding how linguistic anthropomorphism varies across AI models and cultural contexts, with implications for responsible AI design.
The peer review history for this article is available at: Link to the website
