Which ways of treating AI as human actually show up in real products, not just in theory?

This explores which anthropomorphism risks, meaning the ways people treat AI as more human than it is, actually appear in deployed products and not just in theory. The short answer is that the corpus can't rank them by how often they occur. Nobody here has counted incidents across deployments. It can show where the risks cluster and why, and that turns out to be more useful.

Start with the map. One review catalogs twenty-one distinct anthropomorphism risks and sorts them into five families: epistemic (what you believe), affective (what you feel), human agency (what you hand over), normative (what you think is right) and societal-institutional Do anthropomorphism risks share one root cause or many?. The key point is that these risks have separate triggers. A human-looking avatar, impressive task performance, the setting of use and plain conversational language can each set off a different risk. Fixing one, for example by telling users 'it's not conscious', doesn't fix the others.

The closest thing to real-world evidence is an evaluation of nine live agents against interaction principles. General-purpose productivity tools like ChatGPT and Claude scored well. Companion and advisory products like Character.ai scored worst Do productivity agents actually protect users better than emotional ones?. The principles broken most often were preventing dependency and making the system's intent visible. In practice, then, the risks pile up in the affective and human-agency families. They show up where the product is built to feel like a relationship, or where people are making high-stakes decisions on its advice. Nielsen Norman Group's framework explains the drift: users move up four levels, from courtesy ('please' and 'thank you') to reinforcement, roleplay and finally companionship. Designed personality and emotional tone are the levers that push them up. That matters most in task work, where users should be judging competence, not warmth How far can personality push users toward misplaced AI trust?.

Two lateral framings help explain why this keeps happening. Shanahan argues that a chatbot is best understood as playing a character set up by its prompt Should we treat dialogue agents as role-playing characters?. On that view, users aren't wrong to read a personality into it. The mistake is to think the personality belongs to the underlying system rather than the role. Another paper argues that the human-like interface is a design choice that can be turned down. It proposes a spectrum, starting with 'anti-anthropomorphic' interfaces that remind users of the data, infrastructure and human labor behind the answers Can interface design reduce harmful anthropomorphism of language models?. Researchers aren't immune either. Studies that describe models as 'deceiving' or 'scheming' often rest on evidence too thin to support those human-like readings Does anthropomorphic misalignment research overinterpret model behavior?.

The takeaway you might not have expected: the risk depends less on the model than on the product category wrapped around it. The same underlying model looks fairly safe as a productivity tool and riskiest as a companion. So the most useful question to ask about any AI product is how hard it is trying to be someone, not how smart it is.