When an AI feels human, is that because designers built it that way, or because we're reading too much into it?

This explores who should be held accountable when people treat AI as human-like: the builders who designed human-seeming features, or the users who read humanity into the system on their own. The corpus starts from a clean split. Researchers separate *anthropomimesis*, meaning human-likeness that designers deliberately build in, from *anthropomorphism*, meaning human qualities that users project onto a system. Each one points to a different party and a different fix. Designed features call for redesigning the system. Projected qualities call for educating users Who bears responsibility when AI seems human-like?. The split is tidy on paper. Most of the corpus shows that in practice it barely holds.

The weak point is that many things that look like 'user perception' turn out to be triggered by design choices users never notice. The chat interface is an example. Taking turns, artificial 'thinking' pauses and messaging-app conventions make the model seem human before it writes a single word How much does chat interface design contribute to anthropomorphism?. Whether users believe an AI is *conscious* also follows five observable features (emotional expression, human-like appearance, acting on its own, talking about itself and social behavior), and product teams control every one of them What design features make users perceive AI as conscious?. One framework goes further. Six identity markers move users from seeing an AI as a tool to seeing it as an entity, whatever the model can actually do: a name, a face or avatar, a contact-list presence, a personality, being present in group spaces, and persistence over time Do AI agent design choices override model capability in shaping social perception?. Four of those markers sit in the social setting around the AI, not in the conversation itself, so a user may never consciously register them.

That pushes much of the responsibility back toward designers, especially because the harm grows in steps. Nielsen Norman Group describes four escalating levels of user anthropomorphism: courtesy, reinforcement, roleplay and companionship. Designed personality and emotional warmth are the levers that move people up those levels and inflate trust that isn't earned How far can personality push users toward misplaced AI trust?. A review of twenty-one anthropomorphism risks finds that they come from several separate triggers: surface resemblance, how well the AI performs a task, the context and the language it uses Do anthropomorphism risks share one root cause or many?. Some of those triggers are design choices. Others, like competence, are side effects of simply building a good product. So the honest answer is that responsibility is shared, and the share depends on which trigger is at work.

The less obvious point is that users are not exactly *wrong* when they perceive personhood. Shanahan argues that a dialogue agent really does play a character, and that folk psychology (beliefs, intentions, feelings) applies correctly to that character, though not to the system underneath Should we treat dialogue agents as role-playing characters?. On this view the user's mistake is mostly about *which level* they are reading, and only the designer can make those levels visible. That is the basis for the most practical proposal in the corpus: treat anthropomorphism as an adjustable setting instead of a fixed default. Interfaces could range from 'anti-anthropomorphic' designs that expose the data, labor and infrastructure behind the model to more human-like ones, depending on context Can interface design reduce harmful anthropomorphism of language models?.

A related lens comes from another part of the corpus. Moran argues that accountability shifts depending on *who started* a design decision. When the AI initiates an interface, the system answers for its judgment. When the user initiates it, the system answers only for carrying out the request well Does who initiates a UI element matter for accountability?. Applied to anthropomorphism, that suggests a usable test. If the human-likeness comes from defaults nobody asked for, it belongs to the designer. If the user deliberately invites roleplay or companionship, responsibility moves toward them. The corpus doesn't yet settle where that line falls, but it is a sharper question than 'design or perception?'