Can interface design reduce harmful anthropomorphism of language models?

The paper, framed as a design provocation, argues that anthropomorphism should be treated as a design variable rather than accepted as the fixed default of LLM interfaces. It proposes "a spectrum of metaphors from transparency-driven 'anti-anthropomorphism' to uncanny 'hyper-anthropomorphism'" that "introduce materiality to interface metaphors, exposing LLMs as sociotechnical systems shaped by human labor, infrastructure, and data." The paper observes that the anthropomorphic metaphor already dominates current LLM interfaces not only through language but through chat-based conventions — turn-taking and "artificial 'thinking' delays" that mimic human conversation — and that this default goes largely unquestioned even though "metaphors are inherently partial, meant to capture both similarities and differences."

The reasoning is that anthropomorphic cues build trust early on that "snowball into false beliefs about LLM personhood during vulnerable moments," and that masking non-human qualities inhibits users' ability to verify outputs, especially in sensitive or subjective use. Because of this, the paper argues interface metaphors need to actively counteract the default rather than merely hedge language-level claims. It frames the goal as a shift from "optimizing usability" to "encouraging critical engagement," conceding that new metaphors "may complicate user mental models and introduce friction" against standard usability heuristics — but arguing friction is already present in users' lived experience (citing "clanker" as slang reintroduced for AI) and can be protective: "when a user is being pushed to physical or emotional harm due to LLM interactions, friction is not just beneficial, it can be lifesaving."

This reframes anthropomorphism as something designers can dial rather than a fixed property of the interaction, extending Who bears responsibility when AI seems human-like?'s distinction between user-perceived anthropomorphism and designer-built anthropomimesis: the spectrum proposed here is explicitly a designer-side lever meant to push back against unchecked user-side anthropomorphism. The harms cited as motivation — "infidelity," "psychosis," and "suicide" tied to LLM use, and users for whom LLMs "have become intimate and romantic" partners — parallel the distributed-delusion mechanism in How do chatbots enable distributed delusion differently than passive tools?, though this paper locates the point of intervention earlier, in interface conventions rather than conversational stance-taking. It also pairs with How much does chat interface design contribute to anthropomorphism?, which names the specific conventions this spectrum is designed to disrupt.

The paper is an unreviewed arXiv preprint offering a theoretical framework and brief illustrative examples ("visualized in Figure 1"), not a built or tested system. It reports no usability study, prototype evaluation, or measurement of whether anti- or hyper-anthropomorphic metaphors reduce overreliance or harm; its own conclusion calls for future "participatory design workshops," prototypes, and "larger-scale behavioral studies" to test the idea. The spectrum should be read as a design agenda rather than evidence that materiality-based interfaces work.