How Do Language Models Sound Human? A Scoping Review of Anthropomorphic Communication Cues in LLM Output and an Agenda for Cross-Lingual Research

Generative assistants increasingly communicate in ways users experience as human-like, through cues such as first-person self-reference, emotional expression, empathy, and claims of relationship. This scoping review maps how anthropomorphic communication cues in large language model (LLM) output have been conceptualized, operationalized, and studied, and which languages that work covers. Following JBI scoping-review guidance and PRISMA-ScR reporting, I searched OpenAlex, PubMed, the ACL Anthology, IEEE Xplore, and Human–Machine Communication on 29 September 2026 for work from 1 January 2020, and chased citations of four seed studies. Sources were included when they analyzed or conceptualized anthropomorphic cues in LLM-generated text. Of 2,910 records, 69 sources met the inclusion criteria, and 22 reports could not be retrieved. All 69 were charted for study type, system, language, and cue labels. Twenty-three sources state a language, and fourteen more use English examples, prompts, or data without stating it. Where a language is evident, 27 sources are English; the set also includes Korean, Japanese, Chinese, Dutch, African American English, and Queer slang. Thirty-two sources give no evidence of language. English accounts for most sources where a language is visible, but the evidence does not establish that the literature is English-only. Drawing on the charted studies, and on Urdu as an illustrative case, I propose a framework separating language-independent, grammar-dependent, and culture-dependent cues, and outline an agenda for cross-lingual human–machine communication research. Status: Preprint. Manuscript currently under peer review at Human–Machine Communication.