In 1966, Joseph Weizenbaum built a chatbot at MIT called ELIZA. It did almost nothing: it took whatever you typed and reflected it back as a question, the way a parody of a therapist might. Tell it your mother hated you and it would ask why you believed your mother hated you. Weizenbaum's own secretary, who had watched him build the thing and knew exactly how hollow it was, asked him to leave the room so she could talk to it in private. He found that disturbing enough that he spent much of the rest of his career, including a whole book, warning people about what he had made.
Sixty years later the machines are considerably better. Software can pick up the tremor in a voice, the word choices that signal distress, the shift in someone's messaging rhythm that means something is wrong. Some of it already reads emotion more accurately than the average friend does, and it will keep improving, because reading emotion turns out to be a pattern-recognition problem, and pattern recognition is the one thing this technology reliably does. What I want to argue is that none of this touches the part of empathy that ever mattered.
The two halves
Empathy, as we normally use the word, bundles two different things. The first is perception: accurately registering what another person feels. The second is caring: being moved by what you register, and turning toward it when you could have turned away. We talk about them as one skill. They are not, and the machines are currently prying them apart.
The clearest evidence that the perception half automates came in 2023, when researchers published a study in JAMA Internal Medicine comparing physicians' answers to patient questions on a Reddit medical forum with ChatGPT's answers to the same questions. Evaluators rated the chatbot's responses as more empathetic, and it was not close. I would be careful with that finding, though. The raters were judging text on a screen. The physicians were dashing off free replies to strangers, presumably between other things; the model had all the time in the world to be warm. What the study measured was empathic wording, which is sort of my point: wording is the automatable half.
The other half does not automate, because it is not a computation. Caring means the other person's state changes what you do, at some cost to yourself. A friend who abandons their own evening to sit with you is spending something finite, attention they cannot spend twice, and they did not have to. I think nearly all the value of being cared for lives in that 'did not have to.' Being chosen by someone with limited time tells you something about yourself. Being processed, however sensitively, tells you nothing except that you were within range of the sensor.
What it isn't
One clarification, because empathy has a reputation problem. The person who absorbs everyone's feelings until they are drowning in them is not displaying an excess of empathy; they are being flooded, and flooded people are famously useless to the person in front of them. Nurses learn early, out of necessity, that a certain amount of boundary is what makes sustained caring possible at all. People-pleasing is a different failure and needs only a sentence: it is usually about the pleaser's discomfort, not the other person's need.
The strongest objection I know
Here is where I should concede something. In early 2023, the company behind Replika, an AI companion app, abruptly changed how its bots behaved, stripping out the romantic features many users had built their daily lives around. The grief in the user community was real. Moderators of the app's forum ended up pinning suicide-prevention resources. It is easy to be condescending about this, and I would rather not be. Those users were getting something from the machine: comfort, steadiness, a listener at three in the morning, and for some of them it may have been more than they were getting from any human. If simulated care reliably makes lonely people feel better, then my claim that it means nothing is, at minimum, too glib. I am genuinely not sure where the line sits.
But I think the core distinction survives that story. What Replika users lost, judging by their own accounts, was a relationship they had experienced as particular to them. The update revealed that it was not: one policy decision at one company rewrote thousands of relationships at once, identically. That is the whole difference, contained in a single fact. A person's care is theirs to give and theirs to withdraw, one relationship at a time. A machine's care is a product setting.
What gets expensive
So my expectation, and I hold it loosely, is that machine empathy makes human care more valuable rather than less, roughly the way mass-produced furniture made handmade furniture a luxury. When perfectly attuned attention is available to everyone on demand, attunement stops signaling anything. What still signals is cost. The colleague who noticed you were off on Tuesday and followed up on Thursday. The clumsy message from a friend who plainly stopped what they were doing, got a word wrong, and reached out anyway; you can usually feel the difference between that and the beautifully worded one, even if you could never prove it. I build an assessment platform for a living, which biases me toward believing most human qualities can be measured, and even I do not know how you would measure this one. You can score whether someone identified the emotion correctly. Whether they were moved by it does not leave that kind of trace.
The practical consequence is smaller than the argument. When someone tells me something hard, the tempting move now is the well-crafted response, the one a model would draft, and I have started to distrust it, because the well-crafted response has become the cheap one. What is expensive, and I suspect what actually registers, is worse sentences and real attention: the phone in another room, the follow-up question two days later, letting the conversation take the hour it wants to take. None of that is new advice. It is just that the alternative used to be scarce too, and it is not anymore.