How do you tell what you actually designed into an AI's human-like feel, versus what users are just imagining on their own?

This explores how designers can separate the human-likeness they build into an AI (anthropomimesis) from the human-likeness users project onto it (anthropomorphism). The most useful test the corpus offers is about responsibility: ask who would have to change for the effect to go away. If changing the system would remove it, it's anthropomimesis and the fix is redesign. If it would persist across reasonable designs because people bring it with them, it's anthropomorphism and the fix is closer to user education or framing Who bears responsibility when AI seems human-like?. So the distinction isn't just vocabulary. It tells you which of two separate accountability paths you're on.

The catch is that many features that feel like user projection were designed in. Chat interfaces encourage human-likeness before the model writes a single word. Turn-taking, artificial 'thinking' pauses, and messaging-app conventions all cue a social partner, so designers who only audit the model's language miss part of what they've built How much does chat interface design contribute to anthropomorphism?. The same goes for identity choices. Giving an agent a name, a visual identity, a personality, a place in your contacts, a presence alongside other people, and memory that persists can push users to treat it as an entity rather than a tool, whatever the underlying model can actually do. Four of those six markers sit at the social graph level, outside the conversation window where designers usually look Do AI agent design choices override model capability in shaping social perception?. A practical habit follows: list every one of these levers you've set, and treat each as a design decision you own.

That makes human-likeness something you can set to a chosen level, not an on/off default. One line of work proposes a spectrum that runs from deliberately anti-anthropomorphic designs to highly human-like ones. Metaphors drawn from what the system is physically made of can pull users back toward seeing it as it is: something built on human labor, data, and infrastructure Can interface design reduce harmful anthropomorphism of language models?. Shanahan's role-play framing offers a matching mental model. A system prompt sets up a character, and the model produces text consistent with that character. Folk psychology like 'it wants' or 'it believes' fits the character, not the system underneath Should we treat dialogue agents as role-playing characters?. Seen this way, a persona is something the designer authored, and authoring is anthropomimesis.

Knowing which mechanism you're dealing with also tells you which risk you're dealing with. Anthropomorphism risks don't share one root cause. They split into epistemic, emotional, human-agency, normative, and societal harms, triggered separately by surface resemblance, task performance, context, and language Do anthropomorphism risks share one root cause or many?. Some of those triggers, like surface resemblance and interface conventions, are clearly the designer's. Others, like the impression created by strong task performance, emerge whether or not you intended them. The distinction applies to researchers too. Studies that describe models as 'deceiving' or 'scheming' can be committing anthropomorphism in their own interpretations Does anthropomorphic misalignment research overinterpret model behavior?.

A last point you may not expect: making users part of the design process doesn't settle the question. In a study where people co-designed their own preference agents, participants felt well represented, yet independent checks found the agents were more generic than the people's actual responses. The process itself produced a feeling of alignment Does co-design participation hide misalignment in preference agents?. Even 'user-driven' design can quietly build in trust and human-likeness that users then experience as their own perception. The corpus doesn't give designers a step-by-step audit method. What it does give is the responsibility test plus a concrete inventory of levers: interface rhythm, identity markers, persona prompts, and metaphor.