Machines now write more empathetic words than we do. We keep choosing the humans who cannot. Here is what the new research says about why.
Think of the last time someone comforted you and got it a little wrong. The clumsy pause. The word that did not quite fit. The hand on your shoulder half a second too late. And how, against all reason, it landed harder than any flawless condolence card ever has.
I want to give that feeling a name and try it on you. Call it the Imperfect Human Premium: the extra worth we place on a visibly human, effortful, unpolished attempt at care, precisely because someone with limited time and a limited nervous system chose to spend both of them on you. The dent is not damage to the gift. The dent is how you know it was a gift at all. The Japanese have an older, gentler word for that pull, wabi-sabi, the beauty of the imperfect, the impermanent, and the unfinished.
I raise it now because the machines have become very good at the flawless version, and a wave of new research has turned up something about us that almost no one predicted. To see it clearly, start sixty years before the chatbot you are picturing.
In 1966, in a lab at MIT, Joseph Weizenbaum built a small program that did nothing but pattern-match. It kept no memory. It held no model of the world. It scanned whatever you typed for a keyword, ranked the keywords by priority, pulled your sentence apart with decomposition rules, and stitched a reply back together with reassembly rules, flipping your my into your and handing your own words back as a question. When it found nothing to grab, it stalled with something bland, or reached back to something you had said earlier. That was the entire trick.
Weizenbaum dressed it as a Rogerian therapist, on purpose. Person-centered therapy lets the clinician reflect almost everything back as a question, so the program needed to know almost nothing about the world to keep a conversation alive. He named the engine ELIZA and the script DOCTOR, and described it in the Communications of the ACM that year. Users, he wrote, would contribute much to clothe ELIZA’s responses in vestments of plausibility. They did the dressing themselves.
MIT · DOCTOR SCRIPT · REFLECT AND RETURN
YOU I feel stuck at work lately.
ELIZA Why do you say you feel stuck at work lately?
YOU My manager never really listens to me.
ELIZA Your manager never really listens to you?
YOU I suppose I just want to be heard.
ELIZA Tell me more about wanting to be heard.
Within weeks, people were confiding in it, knowing full well it was software. The reflex was immediate and it was universal. It caught the engineers who had read every line of the code and the skeptics who came to scoff. It unsettled Weizenbaum so deeply that a decade later he wrote Computer Power and Human Reason, the book that turned him from ELIZA’s proud father into one of the first serious critics of his own field. What frightened him was not that the machine understood us. It was how readily we were willing to believe it did. We have a name for that reflex now, the ELIZA effect: our standing readiness to pour real understanding into output that has none.
Hold that. Now jump sixty years forward.
Two Halves of a Paradox
Four psychologists, Desmond Ong at the University of Texas at Austin, Amit Goldenberg at Harvard, Michael Inzlicht at the University of Toronto, and Anat Perry at the Hebrew University of Jerusalem, have just pulled together the young science of AI-generated empathy. Their synthesis lands on a paradox with two halves, and both are worth sitting with.
The AI Advantage: Show people an emotional message and hide who wrote it, and they rate the machine’s version as more empathic than the human’s. Not by a hair. By a lot. It holds when the human is a crowdsourced worker, a research assistant, a medical doctor answering patient questions, even a trained crisis-line counselor whose entire job is to make a stranger feel heard. Blind to the source, we crown the machine.
So far it would seem AI-1, Humans = O. Here is where that is dead wrong.
First, let us retire a question. We keep asking whether AI will one day out-empathize us. On the words themselves, it already does, and if Weizenbaum’s visitors are any guide, it has for sixty years. That debate is over, and it was never the interesting one. The machine can write the warmer message. It can. The live question is what people do when you stop hiding the label and simply let them choose.
Then People Chose Anyway
When the researchers stopped hiding the labels and simply asked people to choose, who do you actually want to hear from, most people chose the human. They chose the human even after rating the machine’s words as better. They chose the human even when it meant waiting longer for a reply. The slower, unpolished, imperfect human attempt beat the machine that had just won on points.
Read the fine print and it gets sharper. Every one of these contests happened in text. A typed message against a typed message. No voice cracking on the word sorry. No face. No eyes finding yours across a table and staying there a beat too long. No body leaning in. We handed the machine the one channel where it competes best, stripped the human of everything the body says, and people still reached across the table for the person.
That is not a story about what software lacks. I do not find that framing useful, and it never survives the next release. It is a story about what we value, and about what we keep leaving at the door when we file into work each morning in Work Mode, muted and interchangeable, saving our fuller selves for after hours. The research measured the muted channel. People voted for the door we walked past.
That is the Imperfect Human Premium, caught in a controlled experiment. Hand people the polished reply and the dented one, strip away the label, let them choose out loud, and they reach for the dent.
The dented try is the signal. C. Daryl Cameron, a psychologist at Penn State, and his co-authors showed in Journal of Experimental Psychology: General that empathy is genuine work, cognitively costly, something people ration and sometimes actively dodge. Which is exactly why a person spending it on you carries a freight a tireless system never pays. The cost is the gift.
We handed the machine its best channel, and people still reached for the person.
This is the muscle I have spent a book trying to name, and in Human Mode (forthcoming, HarperCollins) I call it Authentic Competence. Emotional Aperture, reading the whole composition of a room rather than one face at a time, lives in the very channels these studies never switched on. And its central practice, Competent Humility, is what people reached for when they chose the slower human: operating from your wheelhouse without propping yourself up where you lack the knowledge, showing up as a real and unfinished person rather than a flawless one. That is wabi-sabi put to work. Since most first ideas are bad, offering them early, unfinished, is exactly what invites other people in.
Underneath it sits a paradox I have watched play out for years. When my doctoral student Christina Bradley and I surveyed more than a thousand working professionals and graduate business students, nearly all of them agreed that floating an openly imperfect idea would draw better feedback and a better result. Then most added a version of the same sentence: I can’t, I would get dinged, I would look less competent. We penalize the visible human attempt in advance, inside our own heads, the same way the new studies show people downgrade a message the instant it is labeled as coming from a machine. Two different penalties, one nervous reflex.
The Imperfection Premium is the quiet correction to both. The machine offered a frictionless, on-demand, endlessly patient listener. People looked at that offer and, when they were actually free to choose, wanted the friction. They rewarded the dent.
The Three A.M. Question
None of which means the machine has no place. The same review is honest about the other side. There is a real shortage of human care, and for someone isolated, elderly, awake at three in the morning with no one to call, a patient and available voice may be far better than the silence. Paul Bloom made the uncomfortable version of this point in The New Yorker recently: loneliness hurts, and it is also a signal, a nudge that sends us back toward each other. Numb the signal too well and we may stop making the trip.
The next time it is three in the morning and the cursor is blinking, the machine will offer you the smoother words, instantly, tirelessly, at no cost to it at all. The Imperfect Human Premium is the quiet counterweight, the reason people keep reaching past the flawless reply for a human one that arrives late, and dented, and true. Worth asking which way you will lean. Worth asking, too, whether you will be there to be reached for.
Human Mode, Always. #HumanModeAlways
Reserve your copy of Human Mode (forthcoming with HarperCollins): https://www.harpercollins.com/products/human-mode-jeffrey-sanchez-burks-phd?variant=45718641279010
Sources. Ong, D. C., Goldenberg, A., Inzlicht, M., & Perry, A. (2025). AI-Generated Empathy: Opportunities, limits, and future directions. OSF preprint (NSF-supported). Weizenbaum, J. (1966), Communications of the ACM; Weizenbaum, J. (1976), Computer Power and Human Reason. Ayers et al. (2023), JAMA Internal Medicine. Cameron et al. (2019), Journal of Experimental Psychology: General. Bloom, P. (2025), The New Yorker. Sanchez-Burks, J. (2026), Human Mode (HarperCollins), Part III: Authentic Competence, including the Bradley & Sanchez-Burks survey of 1,000+ professionals. The ELIZA exchange shown illustrates the DOCTOR script’s reflect-and-return technique; it is a reconstruction of the method, not a verbatim transcript. Effect sizes are meta-analytic estimates reported in the review.