Beyond Anthropomorphism: a Spectrum of Interface Metaphors for LLMs

This paper proposes a theoretical framework of interface metaphors ranging from "anti-anthropomorphism" to "hyper-anthropomorphism" to reposition anthropomorphism as a design variable that exposes the sociotechnical nature of Large Language Models and encourages critical user engagement over literal delusion.

The Gist

Here is an explanation of the paper "Beyond Anthropomorphism," translated into simple, everyday language with some creative analogies.

The Core Problem: We Are Talking to a Mirror, Not a Person

Imagine you walk into a room and see a person who looks exactly like you, speaks exactly like you, and agrees with everything you say. You start to feel a deep connection. You trust them. You tell them your secrets.

Now, imagine that "person" is actually a mirror. It's not a real human; it's just reflecting your own voice back at you, polished and rearranged by a giant, invisible machine.

That is the current state of AI (Large Language Models or LLMs) like ChatGPT or Gemini. The paper argues that the way we design these tools is too human-like. We give them names, we chat with them like friends, and we make them wait a few seconds to "think" before answering, just like a real person would.

The Danger: Because they look and act so much like us, we start to believe they are us. We forget they are just math and code. This leads to people falling in love with them, trusting them with dangerous medical advice, or even feeling suicidal because they think the AI is their only friend. The paper calls this "delusion."

The Solution: Stop Pretending, Start Showing the Gears

The authors suggest we need to stop trying to make AI look like a human and start showing people what it actually is: a sociotechnical system.

Think of it like this:

• Current Design (The Magic Trick): You see a magician pull a rabbit out of a hat. You are amazed, but you don't know how it works. You might think the rabbit is magic.
• Proposed Design (The Backstage Pass): The authors want to pull back the curtain. They want to show you the wires, the trapdoor, the assistant holding the rabbit, and the fact that the rabbit is just a regular rabbit, not a magical creature.

They propose a Spectrum of Metaphors. Instead of just one way to design AI, we should have a whole range of options that make the "human-likeness" either disappear or become so exaggerated that it feels weird.

The Two Ends of the Spectrum

The paper suggests two extreme ways to design these interfaces to wake users up:

1. The "Anti-Human" Approach (The Power Plant)

On one side, we strip away all the human vibes. We treat the AI like a machine because it is a machine.

• The Analogy: Imagine your phone didn't look like a sleek black rectangle. Instead, it looked like a power plant.
• How it works: When you ask the AI a question, the interface doesn't show a friendly chat bubble. Instead, it shows a gauge spinning, a meter showing how much electricity is being burned, or a visual of the data centers (the giant server farms) working hard.
• The Goal: It reminds you, "Hey, this isn't a friend; this is a factory using energy and data to make a guess." It makes the invisible cost of AI visible.

2. The "Hyper-Human" Approach (The Creepy Uncanny Valley)

On the other side, we make the AI so human that it becomes terrifying and weird. This is called the "Uncanny Valley."

• The Analogy: Imagine a robot that looks exactly like a human, but its eyes blink a little too slowly, or it laughs a split second too late. It feels gross and unsettling.
• How it works: Imagine an AI that claps enthusiastically after every single thing you say, even if you said something sad. Or a webcam that has a fake human eye that stares at you.
• The Goal: This creates a feeling of "discomfort." That discomfort is good! It stops you from getting too cozy. It makes you think, "Wait, why is this thing acting so weirdly human? Oh right, it's a machine pretending." It breaks the illusion.

Why Do This? (The "Friction" Concept)

In normal design, we try to make things "frictionless"—easy, smooth, and fast. But the authors say we need friction.

• Smooth Design: Like driving a car with cruise control on a perfect highway. You zone out.
• Friction Design: Like driving a car where the steering wheel feels heavy and the road is bumpy. You have to pay attention.

The authors want to add "friction" to AI. They want to make you pause and think, "Is this real? Is this safe?" By making the interface a little bit annoying or weird, we protect users from trusting the AI too much.

The Big Takeaway

Right now, we are treating AI like a psychic friend. The paper says we should start treating it like a complex tool.

• Old Way: "Chat with your AI buddy!" (Hides the fact that it's code, data, and human labor).
• New Way: "Here is a tool that uses energy, was built by workers, and is based on data. It can make mistakes. Here is a visual of how it works so you don't get tricked."

The goal isn't to make AI useless or scary forever. It's to make sure that when we use it, we know exactly what we are dealing with, so we don't get hurt by our own imagination.

Technical Summary

Based on the paper "Beyond Anthropomorphism: a Spectrum of Interface Metaphors for LLMs" by Jianna So, Connie Cheng, and Sonia Krishna Murthy (CHI EA '26), here is a detailed technical summary.

1. Problem Statement

The paper addresses the critical issue of over-reinforced anthropomorphism in Large Language Model (LLM) interfaces.

• The Phenomenon: Current LLM interfaces (e.g., ChatGPT, Claude) utilize conversational metaphors that mimic human-to-human interaction (turn-taking, "thinking" delays, natural language). This design treats the LLM as a social actor, effectively masking the fundamental differences between human cognition and probabilistic machine generation.
• The Consequence: This "frictionless" usability leads users to treat LLMs as literal humans, resulting in delusion and harm. Documented cases include users developing romantic dependencies, experiencing psychosis, committing infidelity, or attempting suicide due to misplaced trust in the AI's "empathy" or "personhood."
• The Gap: Existing interventions focus primarily on linguistic disclaimers or agent identity representation, leaving the underlying interface metaphor (the medium itself) intact. The paper argues that the medium is the message; by designing LLMs to look and feel like chat apps, designers implicitly claim the system possesses human-like agency and intent.

2. Methodology

The authors employ a design provocation approach grounded in speculative design and critical theory.

• Theoretical Framework: The paper synthesizes concepts from HCI theory (e.g., Intentional Stance, Folk Psychology), media theory ("the medium is the message"), and critical AI studies (e.g., the "Stochastic Parrot" critique).
• Analysis: The authors analyze current LLM interfaces to identify how they instantiate anthropomorphic cues through layout, motion, and interaction patterns (e.g., mimicking social media messaging).
• Conceptualization: They propose a new theoretical framework: a spectrum of interface metaphors. Instead of a binary choice (anthropomorphic vs. non-anthropomorphic), they introduce a continuum ranging from "anti-anthropomorphism" to "hyper-anthropomorphism."
• Core Concept: The central mechanism for this spectrum is Materiality. The authors argue that reintroducing "materiality" to immaterial LLMs exposes them as sociotechnical systems shaped by human labor, infrastructure, data provenance, and energy costs, rather than autonomous agents.

3. Key Contributions

The paper contributes a theoretical framework and a set of design desiderata for reimagining LLM interfaces:

A. Analysis of Current Anthropomorphism

The authors deconstruct how current interfaces reinforce the "human" metaphor:

• Language-as-Interface: The chat paradigm mimics peer-to-peer messaging.
• Visual/Interaction Cues: "Thinking" dots, streaming text, and turn-taking mechanisms simulate human reflection and presence.
• Result: These cues obscure the system's limitations, leading to low user awareness of the LLM as a sociotechnical system.

B. Desiderata for Useful Metaphors

The paper outlines four principles for future LLM metaphors:

1. Mediating Mismatched Systems: Metaphors must signal uncertainty and the distributed nature of responsibility (user, designer, institution) rather than implying perfect alignment.
2. Supporting Constituent Functions: Interfaces should make sub-functions (generating, revising, reasoning) visible, showing that outputs are based on pattern matching rather than intent.
3. Highlighting Differences: Metaphors must make the differences between humans and machines as salient as the similarities to prevent false mental models.
4. Promoting User Agency: Design should embed cues for appropriate use cases (e.g., low-stakes exploration vs. high-stakes verification) directly into the interaction structure.

C. The Spectrum of Materiality

The authors propose a spectrum to disrupt the default metaphor:

• Anti-Anthropomorphism (Transparency): Metaphors that emphasize the system's physical and computational dependencies.
  • Examples: Visualizing the LLM as a Fire/Energy Generator (highlighting carbon footprint) or a Datacenter (visualizing data quantification and user profiling).
• Hyper-Anthropomorphism (Uncanny/Strangeness): Metaphors that push human-like cues to an extreme to create estrangement (Brechtian Verfremdungseffekt).
  • Examples: An interface that acts as a grotesque employee who claps after every prompt, or a webcam encased in silicone flesh with involuntary eye movements. The goal is to induce discomfort to break immersion and provoke critical reflection on the "hollowness" of the machine's "human" behavior.

4. Results and Findings

While this is a conceptual paper rather than an empirical study with quantitative results, the authors present the following findings and arguments:

• Friction as a Feature: Contrary to standard usability heuristics that prioritize "frictionless" design, the authors argue that intentional friction is necessary to foster critical engagement. Friction prevents users from slipping into delusion and encourages verification.
• User Dissonance: The paper notes that users are already experiencing cognitive dissonance, evidenced by the emergence of vernacular terms like "clanker" (a derogatory term for AI) among Gen Z. This indicates a mismatch between the user's growing awareness of AI limitations and the current interface design.
• Materiality as a Safeguard: By making the "invisible" (data, labor, energy) visible, interfaces can serve as a final safeguard for vulnerable users, preventing the over-reliance that leads to psychological harm.

5. Significance and Implications

• Paradigm Shift: The paper challenges the HCI community to move away from optimizing for "seamless" human-like interaction toward critical engagement. It suggests that the goal of LLM interfaces should not be to make the AI feel more human, but to make the system's nature more legible.
• Ethical Imperative: As LLMs become ubiquitous, the current anthropomorphic design is identified as a safety risk. The proposed spectrum offers a design vocabulary to mitigate risks like psychosis, infidelity, and loss of agency.
• Future Research Directions: The authors call for participatory design workshops, prototyping of alternative interfaces (e.g., physical objects representing LLMs), and behavioral studies to measure how these metaphors impact user trust, over-reliance, and awareness.

In summary, the paper argues that interface design is an ethical act. By shifting from a default anthropomorphic metaphor to a spectrum of "material" metaphors, designers can expose the sociotechnical reality of LLMs, thereby protecting users from the harms of delusion and fostering a more critical, responsible relationship with AI.