Participants were shown a difficult situation, sometimes involving another person and sometimes their own experience, then asked to choose an empathetic response from either a human or artificial intelligence before seeing what either had written. Later, they judged how empathetic the responses seemed and how much effort the writer appeared to have put into them.
The result was a split that goes to the heart of emotional support tools. People rated AI responses as more empathetic and compassionate, and said they were better at making them feel heard. Yet, when choosing whom they wanted to receive empathy from, participants selected the human response about 57% to 58% of the time, a result consistently above chance.
A study published in Communications Psychology called the finding the “AI empathy choice paradox.” The research was led by C. Daryl Cameron, the Sherwin Early Career Professor and associate professor of psychology at Penn State, with doctoral student Joshua Wenger among the researchers.
AI-generated responses also left people feeling more cared for and validated than those from human empathizers. Participants generally saw AI as having put more effort into its answers, and perceived effort predicted both empathy ratings and the feeling of being heard.
That effort mattered differently depending on who was providing it. The link between perceived effort and empathy was stronger for people than for AI, while the difference between human and AI empathy became much smaller when human responses appeared to involve greater effort. Cameron said this suggested people could approach AI-quality responses when they invested more in writing them, though many human responses did not carry that investment.
Even trained crisis responders, compared with AI in one study, fell behind AI on empathic quality in most instances. Still, AI did not reliably surpass high-quality human responses on two measures: a sense of connection to the responder and perceived authenticity.
Cameron and Wenger argued that empathy from people someone knows is built through deep, rich conversations with people who choose to be present. They said AI could not reproduce that quality. Cameron suggested one narrower use for the technology: helping a tired person draft an empathetic message, including a condolence message.
The advantage for AI was larger where the suffering was emotional rather than physical, and larger for compassion than for empathy. A separate study of 21 AI chatbot users identified five components of empathy: understanding, resonating, caring, expressing and helping. Most participants saw resonating and caring as essential features of human empathy that AI empathy lacked.
The researchers also examined what happens when AI takes on more of this emotional work. Four studies found evidence of “empathy offloading”: when people regarded AI as capable of empathy, they became less willing to empathise with other people’s suffering, even without recognising the change. People who learned that someone had received emotional support from large language models engaged less with others’ mental states and offered fewer empathetic responses.
Empathic AI may offer emotional support that is low-cost and easy to access, the researchers said. Their findings also point to a risk that it could weaken the ethical and emotional commitments people make to one another. For developers, a response that wins quality ratings may therefore answer only part of the question.
Comments (0)