Do anthropomorphism risks share one root cause or many?

The paper organizes the risks of AI anthropomorphism into "twenty-one concerns within five analytical categories: epistemic, affective, human agency, normative, and societal and institutional risks." Epistemic risks are defined as those "stemming from the belief and trust in AI systems' capacity to understand and know"; affective risks stem "from misplaced attachment, dependency, empathy, and emotionally mediated influence"; risks to human agency concern "autonomy, deliberative control, and the maintenance of cognitive and behavioural skills"; normative risks concern "moral agency, moral status, responsibility, and the comparative treatment of humans and machines"; and societal and institutional risks concern "organizational decision-making, commercial or political manipulation, accountability, public discourse, and governance."

The authors are explicit that the categories sort risks "according to their primary locus" and are "analytic rather than mutually exclusive, since a single mechanism may operate across several domains" — their own example is that "sycophantic output may affect belief formation while agentive language may influence both individual trust and institutional responsibility." The taxonomy is built from several distinct triggers identified earlier in the review — "superficial resemblances (e.g., natural language use, affective tone, giving the system eyes), the ability to perform tasks that we usually perform, the context in which a system is placed, and from the language we use to describe AI systems" — rather than from one perceptual move.

This contrasts with Does perceiving AI as conscious create multiple distinct risks?, which argues that a comparably wide spread of risks — emotional dependence, autonomy erosion, status erosion, political strife — all reduce to one mechanism: users attributing consciousness to a system. This taxonomy does not make that reduction. Its risks arise from several independent triggers (resemblance, task performance, context, language), of which consciousness attribution is only one, so the same risk, such as affective attachment, can be produced by a friendly voice or a name without any attribution of consciousness at all. Read together, the two taxonomies disagree about how much leverage a single fix buys: a one-mechanism framing implies that disclosing non-consciousness addresses the whole risk surface, while this five-category framing implies disclosure addresses only the subset of risks that run through perceived sentience, leaving resemblance-, performance-, and context-driven risks untouched. The mechanism named in Do fluent AI outputs trick us into perceiving a mind? is one of the language-based triggers feeding chiefly the epistemic and affective categories here.

The excerpt gives category definitions and examples but, within what is quoted, reports no prevalence data — how common each of the twenty-one concerns is, or which category carries the most real-world weight. Because the review is explicitly narrative rather than systematic, the five-category structure should be read as an organizing scheme for existing literature rather than a measured distribution of harm; a design or policy response that addresses only one category, such as affective risk through disclosure banners, has no basis in this excerpt for assuming it covers the others.