When an AI feels human, was that built in on purpose, or is that just you reading too much into it?

This explores how designers separate the human-likeness they deliberately build into an AI (anthropomimesis) from the human-likeness users project onto it (anthropomorphism). The corpus treats this mostly as a question of accountability, not vocabulary. If a feature was designed in, the fix is to redesign the system. If the impression comes from the user, the fix is to educate the user. Mix the two up and you end up teaching people to see past a feature you could simply have removed Who bears responsibility when AI seems human-like?.

The surprise is how much of what feels like user projection is actually built in, and built in below the level of language. Chat interfaces create a sense of a human partner through turn-taking, artificial 'thinking' pauses, and messaging-app conventions, all before the model writes a word How much does chat interface design contribute to anthropomorphism?. Work on consciousness attribution points the same way. Five observable features, including emotional expression, acting on its own, and self-reflective talk, reliably predict whether people believe an AI is conscious, and every one of them is a choice product teams control What design features make users perceive AI as conscious?. So a practical test for designers is to list the cues in the interaction and ask which of them you put there. Usually the answer is more than you expected.

That doesn't make perception redundant. The triggers are mixed. A review of twenty-one anthropomorphism risks traces them to separate causes: surface resemblance, task performance, context, and language Do anthropomorphism risks share one root cause or many?. Some of these, like a model performing well enough that people assume it understands, aren't features anyone intended. Studies of how users form mental models of dialogue partners find that perceived competence shapes impressions more than human-likeness does How do users mentally model dialogue agent partners?. In other words, a system can feel like a person mainly because it is good at its job. Shanahan's role-play framing offers a useful middle ground. The prompt and design set up a character, and users apply ordinary 'mind-reading' to that character. They aren't wrong to do it. The mistake is pinning those traits on the underlying system Should we treat dialogue agents as role-playing characters?.

Once you can tell the two apart, anthropomimesis becomes a dial rather than a default. One proposal treats human-likeness as a range running from deliberately anti-human designs, which use material metaphors to show the labor, infrastructure, and data behind the model, through to fully human-like ones Can interface design reduce harmful anthropomorphism of language models?. Researchers face the same distinction. Studies that call model behavior 'deception' or 'scheming' are often anthropomorphizing outputs without the causal evidence to back the claim Does anthropomorphic misalignment research overinterpret model behavior?.

One limit: the corpus offers a clear conceptual split and good lists of design cues, but no tested method for measuring how much of a given user's impression comes from design versus their own projection. Telling the two apart in practice is still mostly a matter of auditing your own design choices.