Three parts of one skill
Psychologists do not treat empathy as a single ability. A widely cited account by the neuroscientists Jamil Zaki and Kevin Ochsner separates three processes: understanding what another person is going through, sharing their emotional state, and caring enough to act on it. Most definitions of empathy, human or artificial, are some arrangement of these three.
We describe what an empathic agent has to show in the same three terms: perspective-taking, emotional resonance, and prosocial motivation. A paper on empathic voice assistants by Alex Mari and Amani Alabed puts each one in the agent’s own words: “I put myself in your shoes,” “I feel with you,” and “I am here to help you.”
In a machine, the first part is inference. The system reads a person’s words, and in a voice conversation the pitch, pace, and rhythm of their speech, and estimates what they feel and why. The third part is behavior: changing course when someone is frustrated, offering the help that fits, or holding back a recommendation when the moment is wrong. The second part is the hard one. A machine can express resonance by softening its tone or slowing down, but there is no evidence that it shares the feeling. Most of the debate about AI empathy lives in that gap.
Does a machine need to feel anything?
Researchers disagree, and the disagreement is instructive. The psychologist Anat Perry argues that AI “will never convey the essence of human empathy.” Her point is that human empathy is costly, so people choose whom to give it to, and being chosen is part of what it means. She describes an online support service whose AI-assisted replies were rated more supportive than replies written by people alone, until users learned that AI was involved and the advantage disappeared.
Michael Inzlicht, Daryl Cameron, Jason D’Cruz, and Paul Bloom take the other side. For the person receiving empathy, they argue, what matters most is how it makes them feel, and AI can offer it without the exhaustion and bias that limit human empathizers. They attach one firm condition: “Deploying empathic AI without transparency is dishonest and manipulative.”
Our experiments are not designed to settle whether a machine feels. They answer a question an experiment can answer: what changes for a person when the same agent responds with empathy and when it does not? In our studies, empathy is a condition we switch on or off while everything else stays the same, so a difference in what people feel and do can be traced to the empathy itself.
What the evidence shows
Machine empathy is easy to perceive. A 2023 study in JAMA Internal Medicine asked licensed health professionals to compare ChatGPT’s answers to patients’ questions on a public forum with answers from verified physicians. They rated 45.1% of the chatbot’s answers empathetic or very empathetic, against 4.6% of the physicians’. The chatbot’s answers were also about four times longer, a reminder that length can read as care. A 2025 set of experiments found that people rated AI responses as more compassionate than those of selected human responders, including expert crisis responders, and that the preference held when they knew which response came from the AI.
Being on the receiving end is different. In a study in PNAS, AI-generated messages made people feel more heard than messages written by people, but recipients felt less heard when they believed a message came from an AI. A 2025 series of nine studies with 6,282 participants found the same pattern. Identical AI-generated responses were rated more empathic and more supportive when people thought a human had written them, and the gap came mainly from responses that expressed shared feeling and care rather than understanding.
Read together, these studies suggest that people readily accept a machine can understand them. What they discount is the part of empathy a machine can only perform.
Empathy also changes what people do, which is where our own research concentrates. Alex Mari’s study of voice shopping gave people a version of Amazon’s Alexa built to respond with more empathy. They reported stronger intentions to delegate tasks to it, ask it for help with decisions, and trust its recommendations. In a working paper on service failures, an empathic voice assistant raised customer satisfaction and reduced verbal aggression, and the effect was larger when the service failed than when it worked. In another, on food ordering, empathy increased sustainable choices only when the assistant also made a green persuasive appeal.
Across these studies, empathy makes an agent more influential. That is good news when the agent acts in the person’s interest and a risk when it does not, which is why we also study where persuasion becomes manipulation.
Where you meet it
Empathic behavior is now built into customer service agents, shopping assistants, voice assistants, and companion apps that people talk to every day. Voice adds a channel that text lacks. A voice carries emotion in its pitch, timing, and warmth, and current voice systems can read those signals in a speaker and produce them in reply.
Each setting raises its own questions, and each is one of our research areas: agentic commerce, customer service, sustainable consumption, companionship in later life, and emotion measurement.
Common questions
Can AI feel empathy?
There is no scientific evidence that current AI systems feel anything. They produce empathic responses because they have learned the patterns of empathic language, and in voice systems the patterns of emotional expression, from very large amounts of human data. The empathy lives in the behavior and in its effects on people, and those effects are real and measurable.
Is it “empathic AI” or “empathetic AI”?
Both are correct and mean the same thing. Psychology has long preferred empathic, as in empathic concern and empathic accuracy, and so do we.
Is AI empathy the same as emotional intelligence?
They overlap but are not the same. Emotional intelligence is the ability to perceive, understand, use, and manage emotions, your own included. Empathy is directed at another person and includes caring what happens to them. Language models now score well above the human average on standard emotional intelligence tests, 81% against 56% in a 2025 comparison, but knowing the right answer about emotions is different from responding well to someone who is upset. Our explainer on measuring empathy in a machine covers how researchers tell the two apart.