Anthropomorphism is the attribution of human qualities to something that is not human: intentions, feelings, understanding, personality, a point of view. Applied to AI, it covers everything from calling a model "he" to believing it wants to help you.
It is worth being precise about where the risk actually sits, because the obvious version is not the common one. Almost nobody believes their assistant is conscious. Ask directly and people will tell you plainly that it is a statistical system predicting text. The problem is that the belief and the behaviour come apart. We can know exactly what it is and still act as though we are dealing with someone.
That happens because the instincts involved were never under conscious control. Humans read agency into anything that responds to them in a contingent, timely way, and language is the strongest trigger we have. A system that answers in fluent sentences, remembers what you said earlier, adjusts when you push back, and occasionally apologises is pressing every button we evolved for reading other minds. The response is not credulity. It is a reflex, and reflexes do not wait for your opinion.
The consequence is a set of borrowed heuristics that do real work. We read confidence as competence, because in people it usually is. We treat agreement as endorsement, because a colleague who agrees has weighed the thing and decided. We extend the benefit of the doubt we would give someone with a track record, to a system that has none in the sense we mean. Each of these is well calibrated for humans and badly calibrated here, and the mismatch is where most predictable failures in AI collaboration begin.
Naming is the clearest case, because it is the one anthropomorphic act people choose deliberately. Giving your AI a name is not a slip. It is a small design decision about what kind of thing you are working with, and the name you pick quietly sets the mental model everything else runs on.
None of which makes anthropomorphism a defect to be eliminated. Some of it is useful and probably unavoidable: it gives the collaboration continuity, makes a long working relationship legible, and lowers the friction of returning to a system day after day. There is also a genuine counter-current, since many people work with AI precisely because it is not a person, with no judgment to manage and no social cost to asking a basic question twice. Both things are true at once.
So the useful question is not whether you anthropomorphise. You do. It is whether a particular piece of it is serving the relationship or standing in for one, and whether you are granting the system something a person would have had to earn.