Anthropomorphism in the age of Large Language Models: An overview of potential risks and mitigations

Large Language Models (LLMs) and more broadly Artificial Intelligence (AI) systems are often described and understood in human-like terms, a phenomenon known as anthropomorphism. This paper provides a synthesis of recent literature on anthropomorphism in AI, covering theoretical frameworks, the role of language in framing AI as human-like, the various risks of anthropomorphizing machines, and strategies to mitigate these issues. After examining why we tend to anthropomorphize AI systems and whether we are right to do so, we highlight the impact of linguistic framing on anthropomorphism. Then, we introduce a conceptual taxonomy of risks associated with AI anthropomorphism. This taxonomy groups twenty-one concerns within five analytical categories: epistemic, affective, human agency, normative, and societal and institutional risks. Finally, we relate these concerns to proposed interventions in design, communication, education, and governance. We argue that a better understanding of AI systems requires concepts and theories grounded in their organization and demonstrated capacities. The linguistic shaping of anthropomorphic perceptions should form part of this scientific effort, since our descriptions influence both how these systems are understood and the roles we allow them to occupy in society.

Introduction. Anthropomorphism is a natural human tendency to project human-like qualities onto non-human entities [1]. In the context of artificial intelligence (AI), this often means treating AI systems as if they have minds, intentions, or emotions similar to our own. What is particularly different from the anthropomorphism of other non-human entities is that on top of treating AI systems as if they have human-like capacities, we also act on the basis of the suggestions and decisions that they produce, appraising these in ways similar to how we would evaluate those of human counterparts. As AI systems become more sophisticated in what they can do [2, 3], their practical influence over our lives increases. For example, advanced conversational agents can by now produce fluent and contextually appropriate language, leading people to find it "increasingly hard to resist" ascribing human qualities to them [4], treating them effectually as they would a friend or a trusted expert. This is deeply problematic as long as it is still a question whether AI systems actually possess the capacities that we attribute to them [5, 6]. In some contexts, for instance, anthropomorphism can operate as a cognitive illusion that compromises epistemic autonomy by biasing users to interpret system outputs through an unwarranted model of human-like mentality. From this perspective, anthropomorphism is an interpretative stance with epistemic and ethical implications, and therefore, a conceptual analysis is a prerequisite for normative debate [2].

Anthropomorphism of AI systems can stem from 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. Even experts themselves are prone to using intentional language for AI systems, inadvertently "begging the question" by implying the very understanding or agency that is in dispute [4]. Moreover, the anthropomorphic habit of thought is reinforced by decades of science fiction and cultural narratives that portray AI systems as near-human or even super-human intelligences. Yet, human-like performance by an AI system does not necessarily equate to a human-like mind [7, 4].

This article offers a narrative and conceptual review rather than an exhaustive or systematic survey. It focuses primarily on large language model-based conversational systems, while drawing on broader research concerning AI, robotics, human-computer interaction, and mind perception where this helps clarify the relevant mechanisms and concepts. Section 2 examines the cognitive, motivational, cultural, and linguistic factors that make anthropomorphic interpretations appealing, including the distinctive role of natural-language interaction. Section 3 evaluates whether the technical capacities of contemporary AI systems warrant attributions of agency, intentionality, consciousness, knowledge, understanding, reasoning, or moral competence. Section 4 considers how terminology and metaphor mediate the transition from observed performance to claims about underlying capacities. Section 5 organizes the risks of anthropomorphizing AI systems into five overlapping domains, and Section 6 reviews proposed responses to mitigate these risks in design, communication, education, institutional practice, and governance. Finally, Section 7 summarizes the resulting case for calibrated, capacityspecific descriptions of contemporary AI systems.

Related work. Human experience is inherently self-referential: perception, memory, valuation, and action are organized from the organism’s own embodied and affective standpoint. Changeux and Connes argue that human access to reality is mediated by brain-generated models constrained by the physical organization of the brain [8], while Evers describes the resulting epistemic finitude as the condition of being, in a sense, "prisoners of our brains" [9]. A related perspective can be found in von Uexküll’s notion of the Umwelt, according to which organisms encounter their surroundings through speciesspecific perceptual and action capacities that structure what becomes salient and meaningful to them [10]. Human cognition is also self-projective, allowing individuals to imaginatively place themselves in alternative situations and to represent the perspectives of others [11]. Together, these characteristics make one’s own mental life and human mode of experience particularly accessible reference points when interpreting other entities. In situations involving ambiguity, complexity, or social salience, familiar human categories such as beliefs, intentions, desires, and emotions can provide readily available models for making sense of non-human behaviour [12, 13, 14].

2.2 The Specificity of LLMs: How not to Anthropomorphize Machines that Talk?

LLMs disrupt this practical association between linguistic and cognitive capacities. They are among the first technological artifacts capable of generating extended, novel, and contextually appropriate natural-language outputs across a wide range of domains. Because language is ordinarily produced by persons and interpreted as an expression of their beliefs, intentions, emotions, and understanding, such outputs strongly invite anthropomorphic interpretation. Knowing that an interlocutor is a computational system may be insufficient to prevent users from responding to it as though it possessed a human-like mind. We propose the term linguistic pareidolia for this phenomenon: the perception of human-like mentality on the basis of linguistic patterns that resemble the observable products of human thought, even when those patterns do not provide sufficient evidence for the attributed psychological capacities. The analogy with perceptual pareidolia is structural rather than sensory. In visual pareidolia, a configuration of features triggers the perception of a face despite not having been produced by a face; in linguistic pareidolia, patterns such as coherence, self-reference, emotional expression, or apparent responsiveness trigger the perception of an experiencing and understanding subject.

Method. The preceding sections discussed several ways in which human-like terminology, design, and performance can encourage unsupported anthropomorphic interpretations of AI systems. Based on these discussions, we have identified various risks that can be associated with such unsupported anthropomorphic interpretations and organized these risks according to their primary locus:

• Epistemic risks concern the risks stemming from the belief and trust in AI systems’ capacity to understand and know (see Subsection 5.1).

• Affective risks concern the risks stemming from misplaced attachment, dependency, empathy, and emotionally mediated influence (see Subsection 5.2).

• Risks to human agency concern autonomy, deliberative control, and the maintenance of cognitive and behavioural skills (see Subsection 5.3).

• Normative risks concern moral agency, moral status, responsibility, and the comparative treatment of humans and machines (see Subsection 5.4).

• Societal and institutional risks concern organizational decision-making, commercial or political manipulation, accountability, public discourse, and governance (see Subsection 5.5).

These categories are analytic rather than mutually exclusive, since a single mechanism may operate across several domains. For instance, sycophantic output may affect belief formation while agentive language may influence both individual trust and institutional responsibility. The taxonomy identifies the primary locus of each concern without treating the categories as causally independent or exhaustive. In what comes next, we review the risks within each of those categories, presented in Figure 1.

Discussion. 5.1 Epistemic Risks One major risk is the overestimation of AI systems’ capacities and the corresponding miscalibrated trust (e.g., reliance) in AI systems’ outputs. Anthropomorphic cues like a human name, conversational style, or a friendly voice, may encourage users to respond to an AI system as if it were a human interlocutor and to grant its statements credibility or authority that is not warranted by its actual technical capacities [131]. This risk is especially salient when AI-generated outputs use confident and personal expressions, since linguistic fluency and socially appropriate responses can be mistaken for factual accuracy or domain expertise. Still, the relationship between anthropomorphism and trust is not straightforward. In a preregistered study, Colombatto, Birch, and Fleming found that participants who attributed greater intelligence-related capacities (such as knowledge, memory, and reasoning) to ChatGPT were more likely to accept its advice [132]. By contrast, attributions of experience-related capacities, such as emotions and subjective feelings, were weakly associated with less advice-taking, while attributing consciousness to the system showed no positive association with advice acceptance [132]. These findings suggest that different forms of mental-state attribution affect trust differently: perceived competence appears more relevant to advice-taking than perceived sentience or emotional experience. Although the study does not establish that these attributions caused participants to accept the advice, it shows that attributing knowledge or intelligence to an AI system can accompany increased reliance on its outputs, including potentially inaccurate ones.

An especially consequential form of such reliance is what Shaw and Nave call cognitive surrender: adopting an AI-generated answer with minimal scrutiny, allowing it to override one’s own intuitive or deliberative judgment [133]. In three preregistered experiments involving 1,372 participants, the authors gave participants the option of consulting an AI assistant while completing an adapted Cognitive Reflection Test and experimentally varied the accuracy of the AI system’s recommendations. Access to accurate AI advice improved performance, whereas faulty advice reduced accuracy relative to a condition without AI assistance; consulting the AI system also increased confidence, even when its recommendations led participants to incorrect answers [133]. Cognitive surrender should therefore be distinguished from cognitive offloading more generally. Delegating part of a task to an external tool can be beneficial when the user retains evaluative control. Cognitive surrender happens when the external output supersedes rather than supports the user’s judgment. Shaw and Nave did not test whether anthropomorphic cues cause this behaviour. Nevertheless, their concept identifies a plausible epistemic consequence of framing AI systems as competent agents or authoritative interlocutors: users may cease to treat their outputs as claims requiring evaluation and instead defer to them as judgments.

5.2 Affective Risks Anthropomorphic and related forms of social framing can lead to emotional over-attachment and misplaced empathy. When an AI system is given a human voice, face, or apparent personality, users may begin to perceive and treat it as a social counterpart, potentially attributing to it feelings and needs. It is important to note here that responses are not exclusive to anthropomorphic framing; they may also arise through sociomorphing, whereby an entity is perceived in social terms without necessarily being attributed specifically human characteristics, or through animism, whereby a non-living system is ascribed sentience, subjective experience, or an inner life.

Conclusion. The increasingly human-like language and performance displayed by contemporary AI systems make questions about their underlying capacities difficult to avoid. Yet these impressions do not establish that the systems possess the human-like properties some users may attribute to them. Anthropomorphic impressions arise through an interaction among the system’s outputs, linguistic framing, interface design, users’ cognitive and social dispositions, and the institutional contexts in which these technologies are presented.

The central conclusion of this review is that descriptions of contemporary AI systems should distinguish their functional achievements from unsupported claims about consciousness, intentionality, understanding, or moral agency [6, 76]. Fluent and context-sensitive performance is evidence of what a system can produce under particular conditions, while leaving questions about its underlying capacities unresolved. This leaves room for disagreement about artificial mentality, including whether artificial systems might instantiate forms of mentality that differ substantially from our own, while avoiding conclusions that go beyond the available evidence. The conceptual distinctions developed throughout this review are intended to support this separation. The notion of linguistic pareidolia captures how human-like linguistic patterns can evoke the impression of an understanding or experiencing subject despite limited evidence about the capacities underlying them, while the notion of the agentic tool provides a way of describing increasingly capable and action-oriented AI systems without automatically equating operational autonomy with human-like intentionality or mentality. The usefulness of these terms, like that of other proposed terminology, ultimately depends on whether they allow for more accurate interpretation and clearer reasoning about what these systems are and how they should be used.