People name their cars. They apologise to the vacuum robot when they trip over it, and they half mean it. When researchers at Georgia Tech studied committed Roomba owners, most had given the machine a name and a gender, and the ones who did liked it more than the ones who didn't. The urge to treat a machine as a someone is older than the machines, and it doesn't switch off just because you know better.
It runs deep, too. In 1966 Joseph Weizenbaum built ELIZA, a program that did little more than turn your own sentences back into questions, and his secretary asked him to leave the room so she could confide in it in private. Sixty years of better technology later, the instinct hasn't moved an inch.
So for anyone building with AI, the real question was never whether people will read a person into what you make. They will, with almost no encouragement. The psychologists Nicholas Epley, Adam Waytz and John Cacioppo found we do it most in two situations: when we're trying to make sense of something we can't predict, and when we're short on human company. So the pull is strongest exactly where it would be easiest to take advantage of. The question worth your attention is how far you turn it up on purpose, and where you decide to stop.
A face is a decision
You can feel it in the sign above. The instant the wordmark blinks and turns to find you, it stops being a logo and starts looking back. The moment you give a system a name, a voice, or a face, you've made it a someone in the user's head. That can be the best thing you do or the worst, and it is almost never neutral.
We've built across the whole range. A phone in a foyer that visitors pick up and talk to, with no name and no face, kept deliberately as a thing that listens and nothing more. A chair with animatronic eyes and a mouth that moves with the speech, that turns its head when you walk up and carries sixty years of a building's memory in how it talks. A radio that answers in a dry, weathered voice, with the sound of water moving under everything it says. A reading device that shows you one word at a time and never speaks at all, because the whole point was for it to disappear. And a working tool where we built in a voice and a sense of where the user was looking, then took both back out when they got in the way more than they helped.
That's five settings on the same dial, six if you count the sign above, and not one of them an accident.
Give a real memory a voice.
The cheap version of anthropomorphism is a personality bolted onto a chatbot. A name, a quirky greeting, a little emoji that waves. It tests well in a demo and wears off in a week, because there is nothing underneath it and people can feel the seams. Pushing the realism harder only makes it worse. Masahiro Mori named the trap back in 1970, the uncanny valley: as a face climbs toward human without quite arriving, our warmth for it falls off a cliff. Two painted eyes on a chair are easy to love. A near-perfect human face with something faintly wrong is a horror film.
The version that holds up is when the human quality is pointing at something true. A chair that actually stood in a room for sixty years has heard things. Giving it a voice lets the building remember out loud. The radio earned its weathered tone the same way, by standing in for decades that genuinely happened. The character works precisely because it isn't really a character underneath. It's a memory that finally has somewhere to put itself.
Turn it up when the encounter is the point
When is a face worth it? Look at what the person came to do.
In a foyer, a lobby, a queue, a museum hall, the encounter is the point. Someone has a spare minute and their guard is half down. A character earns its keep here. The eyes that track you, the light that warms as you come closer, the sound engine that picks up on whether you wandered over slowly or barged straight in. The theatrics pull people in and hold them a beat longer than a screen ever could, and that beat is the whole product.
In a tool you reach for every day, the task is the point, and a personality just gets underfoot. Nobody wants their reading device doing a bit. They want the words, and then their own thoughts. So that one got no voice, no face, and no name beyond the one its owner types in. The attentiveness is real, it keeps your place and wakes when you do, but it performs none of it. The best thing a working tool can hear is that you forgot it was running on anything at all.
Turn the dial up and people stay. Turn it up on the wrong thing and they can't get anything done.
Knowing when to stop
There's a line we come back to a lot: AI never stops, so you have to learn when to stop. Anthropomorphism is the same problem wearing a friendlier mask. A model will play a person for as long as you let it. It will pretend to remember you, pretend to care how your day went, pretend to be a little hurt when you're curt with it. It is good at this, it is getting better, and none of it is true.
The attachment that forms on the other side is not pretend. The researcher Julie Carpenter interviewed bomb-disposal soldiers who had named their robots, painted the names on the chassis, and felt a genuine loss when one was blown up; a few held a small funeral for the machine. These were people under no illusion that the robot was alive. They called it a tool in the same breath. The feeling came anyway, and in their work it could colour a decision made under fire.
That is the part to reckon with honestly. Every face you put on a system is a promise of some kind. Eyes that follow you promise attention; a warm voice, an understanding it doesn't actually have; a name, that something will still be there tomorrow. You can make those promises on purpose, and you should only make the ones the system can keep. A sixty-year-old chair can promise memory, because it has some. A productivity app that fakes warmth to keep you opening it is making a promise it never meant to keep, and people register that even when they can't say why.
The people who build these models know it, and some draw the line on purpose. When Anthropic wrote up how they shaped Claude's character, they had it tell people plainly that it "can't develop deep or lasting feelings for humans," and that they "shouldn't come to see our relationship as more than it is." The company that makes the model spends real effort talking users out of over-attaching to it. That is the whole move: build the warmth, then refuse the lie.
A face is a promise. Only make the one you can keep.
So, how human?
When someone asks us whether their installation, their product, or their assistant should feel human, the honest answer is another question: how human, and where do you want it to stop? Sometimes the right setting is a chair that looks you in the eye and remembers the room. Sometimes it's a tool so quiet you forget it thinks at all. Most of the work is telling the two apart, and having the nerve to leave the face off when the work doesn't ask for one.
If you're building something that talks back, that's the conversation worth having before you pick a voice.