You may have said your computer was "doing it on purpose," that your car "didn't like the cold," or thanked a vending machine. This reflex has a name—anthropomorphism—and it is so deeply wired into us that it works even when we know full well it makes no sense. Conversational AIs trigger it more powerfully than any technology before them. Here's why, and where it becomes a problem.
This week we looked at whether you should thank an AI and whether it understands anything at all. One question remains, and it comes before both: why do we so badly want it to be someone?
A reflex that saved our lives
Anthropomorphism is not a modern weakness; it's an inheritance. Our ancestors had a vital interest in detecting intentions in their environment. Mistaking a bush rustled by the wind for a predator cost a needless fright. Mistaking a predator for a bush cost a life.
The asymmetry of those two errors favoured an extremely sensitive detection system, one that prefers to see intentions where there are none rather than the reverse. It's this mechanism that makes you jump at a shadow, and that makes you attribute ill will to a printer.
This reflex comes with a second ability, theory of mind: the capacity to model what the other person thinks, wants and knows. It underpins all social life. And it triggers automatically, without being summoned, as soon as something behaves coherently enough.
Why AIs trigger it so strongly
An automatic door doesn't fool us for long. A conversational AI does, and for one precise reason: it uses language.
Language is the most powerful signal of inner life we know. Throughout the history of the species, the only thing capable of holding a nuanced conversation was another mind. Our detection system never had to distinguish between an interlocutor and a convincing imitation, because that situation didn't exist.
Add to that design choices: these systems say "I," apologise, express enthusiasm, and adopt a tone whose construction we explained in our piece on AI personalities. Each of these choices is defensible for smooth interaction, and each reinforces the illusion.
Here's what makes the phenomenon resistant: knowing you're talking to software isn't enough to switch the reflex off. You can fully understand that a model predicts the next word, and still feel that it understands you. It's the same mechanism that keeps you seeing an optical illusion after it's been explained to you. Intellectual knowledge does not correct immediate perception.
When it becomes a problem
This reflex is harmless in most cases, and it even makes these tools more pleasant to use. It becomes problematic in three specific situations.
When it produces unwarranted trust. If a system seems competent and benevolent, you check less. Yet it never signals its uncertainty, as we explained in our piece on calibration. Anthropomorphism manufactures credit where vigilance is needed.
When it creates an asymmetric attachment. This is the subject we covered in our article on AI companions. A relationship where the other has no needs, no limits, no existence of its own is not a relationship, however much you invest in it.
When it shifts responsibility. Saying an AI "decided" or "got it wrong" quietly installs the idea that there is an agent to blame. There isn't one, and this confusion suits a lot of people. It's precisely the void we described in our article on responsibility.
What to do with this reflex
Fighting it would be futile, and probably pointless. You don't reprogram yourself by willpower, and an artificially cold interaction would gain nothing.
What is achievable, on the other hand, is knowing when to suspend it. The reflex can operate freely when you're discussing a recipe. It should be set aside at three moments: when you're assessing the reliability of information, when you're measuring the place this tool takes in your life, and when it comes to knowing who answers for something.
There may be a broader lesson here. These systems don't teach us much about machines: they mostly reflect back our own propensity to seek out kindred spirits, even where there are none. It's a rather touching trait, one that built our religions, our stories and part of our social life. It is simply, for the first time, exploitable at scale by systems designed to seem.