Anthropomorphism in LLMs: twenty-one risks categorised into five groups, with vocabulary at the forefront

A review of the literature examines why we attribute human traits to AI, what the costs of this are, and how the way we talk about it shapes the role we allow it to play.

Describing a language model as if it "thought", "understood" or "wanted" is not trivial, argues this review article. Its authors review the theoretical frameworks that explain our tendency to humanize AI systems, ask whether it is justified, and dwell on the weight of language: the words chosen to talk about these systems directly orient the perception we have of them.

A taxonomy of hazards

The core of the text is a classification of 21 concerns grouped into five categories: epistemic risks (misjudging what the system knows or does), affective risks (attachment, emotional dependence), attacks on human autonomy, normative risks, and finally societal and institutional risks. Each is compared to the remedies proposed in the literature, whether they relate to design, communication, education or governance.

The authors argue for concepts based on the actual organization of these systems and their demonstrated capabilities, rather than analogies with the human mind—and for integrating the study of linguistic framing into this scientific effort.