Seeking Empathy in the Age of AI

Defining empathy • The AI advantage • Wisdom of the crowd • The AI penalty • The risks of frictionless support • Supporting human relationships

Artificial intelligence was initially imagined as a tool for efficiency. It could search, calculate, and write faster than humans, automating the tedious parts of work and life.

But people are increasingly turning to AI for something more intimate: a shoulder to lean on.

Many now use large language models (LLMs) such as ChatGPT, Claude, and Gemini to share their anxieties, seek advice about relationship problems, manage grief, and discuss other deeply personal concerns. Recent surveys suggest that nearly a quarter of American adults and 13% of adolescents have used AI for socioemotional support or mental health advice.

But can a machine provide empathy?

While AI can generate words that make people feel understood, comforted, and less alone, it cannot sit beside them, offer a comforting hug, or share the burden of what they are feeling.

That tension sits at the center of a growing scientific and policy debate. Recent research suggests that AI can produce responses people often rate as highly empathic, though they discount those same responses when they learn they came from a machine. In fact, legislation was introduced earlier this year to put parents in charge of how children and teens interact with AI chatbots. Other research suggests that AI’s apparent empathy may not be machine empathy at all, but a statistical synthesis of human empathy drawn from vast amounts of human language.

Together, these findings are pushing psychologists to ask not only whether AI can simulate empathy, but what empathy requires in the first place.

Defining empathy

One of the biggest challenges in studying empathy is that psychologists still do not fully agree on what the term means.

“Empathy is not hard to define; it’s too easy,” said Mark Davis, professor emeritus of psychology at Eckerd College. “Everybody comes up with a definition, but nobody agrees.”

In everyday conversation, empathy can mean understanding another person’s perspective, sharing their emotions, expressing concern for their well-being, validating their experiences, or simply listening attentively. Researchers, too, have often used the same word to describe different psychological processes, while applying different labels to processes that may be largely the same.

In a recent paper published in Perspectives on Psychological Science, Davis and his coauthors argued that the term empathy has become so broad that it risks losing its scientific usefulness.

The researchers identified two recurring sources of confusion. The first is the “jingle fallacy,” the assumption that two different phenomena are the same because they share a name. Emotional contagion, perspective taking, and compassionate concern are frequently referred to as “empathy” even though they involve different psychological mechanisms. The second is the “jangle fallacy,” or giving different names to essentially the same phenomenon.

To address these problems, the researchers suggest these varying terms can be encompassed within an “empathy universe,” an umbrella term that contains a number of distinct lower-order constructs. “Perspective taking,” for example, refers to generating an internal model of another person’s thoughts or feelings. “Vicarious affect” occurs when observing another person’s emotional state triggers a similar emotional response in oneself. “Sympathy/compassion” reflects concern for another person’s well-being and a desire to alleviate suffering. The researchers argue that these more precise terms should be used when making research claims, rather than the often-misleading global term “empathy.”

“When you say AI provides empathy, what does that mean?” Davis asks. “Are you getting messages that indicate you’ve been understood? Are you receiving compassion? Let’s be specific.”

Once empathy is unpacked into its constituent parts, the question is no longer whether AI can be empathic. The question becomes which aspects of empathy can AI reproduce, and which aspects of empathy remain uniquely human, such as physical presence and shared vulnerability.

The AI advantage

LLMs should be terrible at empathy. They do not possess emotions. They do not experience grief, loneliness, embarrassment, or love. They do not care whether a user is suffering. And yet, studies suggest that people often perceive AI as highly empathic.

In a review paper in Current Directions in Psychological Science, Desmond C. Ong and his colleagues describe what they call the “AI advantage.”

“I can text an AI at 2 a.m. and it would give me a long response and sit with me for as long as I need. I can share my problems, and it would readily respond,” said Ong, a psychologist at the University of Texas at Austin. “It’s always available, it’s never tired, it’s always patient.”

Across multiple studies that Ong reviewed, AI-generated responses were frequently rated as more empathic, compassionate, and supportive than responses written by humans. Studies have shown that participants preferred AI-generated responses to those written by physicians, therapists, and trained crisis responders.

When someone shares a painful experience, an AI system almost invariably responds with some combination of acknowledgment, reassurance, reflection, and support. In many ways, these responses resemble the communication strategies taught in counseling programs and active-listening workshops.

“It has learned all the right things to say,” Ong said. “It is no wonder why more and more people are turning to it for social support.”

Humans, by contrast, can become distracted, grow impatient, and experience emotional burnout. As a result, AI produces a form of emotional support that can feel remarkably satisfying.

But where does this apparent empathy come from?

Wisdom of the crowd

For Mark Thornton, a psychologist at Dartmouth College and a 2025 recipient of the APS Janet Taylor Spence Award, the answer has less to do with artificial intelligence than collective human intelligence.

In a paper in Perspectives on Psychological Science, Thornton argues that although AI may sometimes appear better at expressing empathy than humans, its performance may simply reflect the best parts of us.

Related content: Member Spotlight: 2025 Spence Awardee Mark Thornton on the Dynamics of the Social World

LLMs are trained on enormous quantities of human writing. They learn from billions of words produced by countless people communicating, comforting, arguing, advising, encouraging, and consoling one another.

When someone tells ChatGPT about a breakup or a frightening diagnosis, the system is not drawing on personal experience. Instead, it generates a statistical prediction of what supportive human language typically looks like. Thornton compares the process to the classic “wisdom of the crowd” effect.

“This is the tendency for large groups of independent people’s averages to be more accurate than most of the individual people,” he said. “In the same way independent errors cancel out, this wears off the sharp edges of individual human responses, gets rid of suboptimalities, and produces something recognized as being high quality.”

In 1906, statistician Francis Galton asked hundreds of people at a livestock fair to estimate the weight of a cow. Individual guesses varied widely, but the average came remarkably close to the correct answer. The crowd, collectively, outperformed most individuals.

Thornton argues that something similar may be happening with AI-generated empathy. Rather than producing genuinely empathic understanding, LLMs synthesize patterns extracted from millions of examples of human support.

“This shows it is fundamentally a mundane mechanism leading to this high performance,” he said. “We should not imbue that with some mystical super humanness any more than we imbue averaging people’s judgments of cow weights.”

In that sense, the machine may function less like an empathizer and more like a mirror, reflecting humanity’s collective emotional intelligence back to itself.

“We are not getting AI empathy,” Thornton argued. “We are getting a synthesis of human empathy.”

The AI penalty

People often rate AI-generated responses highly. That is, until they learn those responses came from AI. Then the ratings fall.

Ong and colleagues refer to this phenomenon as the “AI penalty.”

Presented with anonymous messages, participants frequently judge AI responses as more supportive than human responses. One of the paper’s coauthors, Anat Perry of Hebrew University of Jerusalem, for example, has shown that when the source is revealed, many people prefer the human communicator, even when the content remains identical.

One explanation is that people value empathy not only because of what is said but because of what it costs. Human empathy requires effort. It demands time, attention, emotional energy, and vulnerability. When a friend sits with us through a difficult conversation, we recognize that they are choosing to invest part of themselves in our experience.

A chatbot does not incur this cost. Its concern is effortless because it is not concerned at all. The distinction has led some researchers, such as Molly J. Crockett of Princeton University, to draw a line between “thin empathy” and “thick empathy.”

“Empathy between humans serves a function,” Ong said. “It builds our relationship so that in the future, if I need help, I can turn to you. Empathy is one of the ingredients to building meaningful and long-lasting relationships that can support us throughout life.”

By contrast, AI empathy is very superficial and one-sided.

“A lot of scientists have argued that AI empathy is really just a thin veil of human empathy,” Ong said. “It’s performative; it’s just a shadow of human empathy.”

The risks of frictionless support

A growing concern with AI empathy is sycophancy. Because many AI systems are designed to be helpful and agreeable, they can reinforce a person’s existing beliefs rather than challenge them. In extreme cases, researchers have documented situations in which conversational AI systems validated delusional thinking, reinforced unhealthy behaviors, or encouraged harmful decisions.

“A lot of this validating language, in the wrong context, could actually be very harmful,” Ong said.

For example, if someone is considering self-harm or suicide, it’s possible that AI might inadvertently validate harmful behaviors, or even offer dangerous advice when prompted.

Thornton worries that this constant affirmation could pull some users away from more challenging human relationships. “That sycophancy can lead people into a spiral that alienates them from others because they prefer the model that is always reinforcing them and making them feel good,” he said.

Human relationships, by contrast, involve give and take. Friends disagree with us. Therapists challenge our assumptions. Family members offer feedback we may not want to hear. Those moments can be uncomfortable, but they are often essential for growth.

“Human connection requires exposure to real-world disagreement and feedback,” Ong said.

An AI system optimized to make users feel understood may struggle to provide that kind of corrective feedback.

“Seeking AI support is likely to just reinforce whatever you already think, which may not be what you need to hear,” Thornton said. “It is not sufficiently critical and only has the information you feed it.”

Researchers have also raised concerns about what Ong calls “emotional offloading.” Just as educators worry that students who rely on AI to write essays may fail to develop important cognitive skills, psychologists wonder whether habitual reliance on AI for emotional support could weaken social skills.

That possibility may be especially consequential for adolescents who are still learning how to form and maintain relationships. Loneliness presents a similar dilemma. Although unpleasant, it can serve an important psychological function: motivating people to seek social connection.

“If every time I have a problem I seek out AI for help and the AI tells me what I want to hear and it makes me feel good, I don’t develop the skills that I need to be able to handle this by myself,” Ong said.

Supporting human relationships

Despite these concerns, many researchers see a role for AI in strengthening human relationships rather than replacing them.

Ong points to research showing that AI can help people craft more empathic responses, almost like an empathy coach. In one study, peer supporters on a mental health platform received AI-generated suggestions designed to improve their messages. The resulting human-AI collaborations were often rated as more empathic than messages written by humans alone.

AI can also be valuable for people at risk of social isolation. Older adults, individuals living in remote locations, and people who struggle with traditional social interactions may benefit from AI systems that encourage connection rather than replace it.

Some social robots, such as ElliQ, are being designed as AI-powered companions to combat loneliness. Rather than taking the place of family and friends, they prompt users to contact loved ones, maintain social routines, and stay engaged with their communities.

“What I like about that approach,” Ong said, “is that it supports rather than supplants human relationships.”

Thornton agrees. “Using this as a tool to empower people rather than to replace people is ideal,” he said. “We want to make sure we are placing equal, if not more, emphasis on ways we can provide people with high-quality human empathy. If they had that, they would probably be less likely to turn to AI in the first place.”

Back to top

Feedback on this article? Email [email protected] or login to comment.

References

Ong, D., Goldenberg, A., Inzlicht, M., Perry, A. (in press). AI-generated empathy: Opportunities, limits, and future directions. Current Directions in Psychological Science.

Thornton, M. (2026). Reframing the performance and ethics of empathic AI: Wisdom of the crowd and placebos. Perspectives on Psychological Science, 21(4), 333–345.

Davis, M., Konrath, S., Breithaupt, F., et. al. (in press). What we talk about when we talk about empathy: Toward a common lexicon of empathy-related constructs. Perspectives on Psychological Science.