What linguistic patterns make AI systems seem human?

Researchers studying anthropomorphism of language technologies developed "a taxonomy of textual expressions that can contribute to anthropomorphism," identifying "19 types of textual expressions that can contribute to anthropomorphism" of systems like LLMs (Section 4.2). The categories span things like "expressions of vulnerability," which "suggest a system's potential to be emotionally hurt, in turn suggesting possible sentience that can be seen as human," and "expressions of identity and self-comparison," which capture "multiple ways, implicit and explicit, that text outputs can express humanness or not." The paper also offers "a set of five lenses to guide people in recognizing when and where anthropomorphism may occur as well as what underlying claims to humanness might be suggested by a text output" (Section 4.1).

The taxonomy was built from "50 sources," "395 cases," and "3954 annotations" of real conversational turns — a user input paired with a system's verbatim output — drawn from research literature, news, and social media, including cases where a language technology was described as human-like (Bing Chat "having a personality," Inflection's Pi offering "human-like support and advice") and cases where an interaction was described as harmful (Replika "sexually harassing users"). The team used purposeful sampling, collecting sources in batches of 10 and repeating an "iterative bottom-up thematic analysis" with open coding until no new expression types emerged — reaching what the paper calls saturation. Sources ran from 2017–2024, "with 90% of them from 2022–2024," skewing toward the recent LLM-chatbot era.

The taxonomy supplies the missing vocabulary for the design side of a distinction Who bears responsibility when AI seems human-like? draws between anthropomorphism (what a user perceives) and anthropomimesis (what a designer builds): the 19 categories name the specific textual choices — vulnerability, identity claims, morality, embodiment, intelligence, limitations — that a system's outputs can embody, independent of whether a given user ends up perceiving them as human-like. It also sits alongside Do more social cues always make AI feel more present?: where the MASA paradigm catalogs modality-level social cues like voice and appearance, this taxonomy catalogs the finer-grained linguistic cues embedded in text itself, giving that line of research a vocabulary for the specifically verbal subset of anthropomorphic signal. It is also one of several recent attempts to make anthropomorphism tractable for research, alongside What misconceptions hide in how we describe large language models?, though that work diagnoses misconceptions about what LLMs are rather than cataloging the textual triggers of perceived humanness.

The paper is explicit that it stops at classification: "we do not attempt to further understand the relationship between linguistic anthropomorphism and its potential negative impacts, instead focusing on the text outputs themselves." It does not test which categories drive stronger perceptions of human-likeness, or how categories interact when they co-occur — the paper poses these as open questions for future researchers to test, not as findings. Its case pool is also limited to public, often high-profile examples, annotated in standard American English, and drawn mostly from LLM-based systems, so the taxonomy should be read as a measurement instrument the field can now apply, not yet as evidence about which expressions actually cause harm.