Anthropomorphism represents a major challenge in human-chatbot interaction, as users often attribute human characteristics to conversational agents despite their purely simulated nature. The issue is particularly relevant for Large Language Models (LLMs), whose advanced conversational abilities closely resemble human communication, reinforcing attributions of human-like traits. Identifying anthropomorphic language in responses generated by LLM-based chatbots is important because it can significantly influence user experience while also fostering inappropriate trust, potentially harmful decisions, and excessive reliance on these systems. In this paper, we investigate whether LLMs can recognize anthropomorphic language in responses generated by LLM-based chatbots by comparing human annotations with LLM judgments under two conditions: one guided by an existing taxonomy of anthropomorphic linguistic expressions and one without such guidance. By analyzing 1,000 real-world conversational turns from five different LLM-based chatbots, we find that explicit guidance improves precision but does not improve recall for subtler stylistic and paralinguistic markers of humanness, which remain a blind spot for both guided and unguided LLM judges. We also find substantial variation across the five LLM-based chatbots in the degree to which anthropomorphic language is expressed and identify qualitative differences in how anthropomorphism manifests across them. Based on these findings, we discuss open research questions surrounding anthropomorphism in LLM-based chatbots and its automatic detection.
Can LLMs Tell When They Sound Human? A Study of Guided and Unguided LLM Judgments of Anthropomorphic Language
Villata S.
First
;Bonarelli A.;Di Caro L.;Rapp A.Last
2026-01-01
Abstract
Anthropomorphism represents a major challenge in human-chatbot interaction, as users often attribute human characteristics to conversational agents despite their purely simulated nature. The issue is particularly relevant for Large Language Models (LLMs), whose advanced conversational abilities closely resemble human communication, reinforcing attributions of human-like traits. Identifying anthropomorphic language in responses generated by LLM-based chatbots is important because it can significantly influence user experience while also fostering inappropriate trust, potentially harmful decisions, and excessive reliance on these systems. In this paper, we investigate whether LLMs can recognize anthropomorphic language in responses generated by LLM-based chatbots by comparing human annotations with LLM judgments under two conditions: one guided by an existing taxonomy of anthropomorphic linguistic expressions and one without such guidance. By analyzing 1,000 real-world conversational turns from five different LLM-based chatbots, we find that explicit guidance improves precision but does not improve recall for subtler stylistic and paralinguistic markers of humanness, which remain a blind spot for both guided and unguided LLM judges. We also find substantial variation across the five LLM-based chatbots in the degree to which anthropomorphic language is expressed and identify qualitative differences in how anthropomorphism manifests across them. Based on these findings, we discuss open research questions surrounding anthropomorphism in LLM-based chatbots and its automatic detection.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026-AIxIA.pdf
Accesso aperto
Tipo di file:
PDF EDITORIALE
Dimensione
193.94 kB
Formato
Adobe PDF
|
193.94 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.



