New research suggests that empathy coming from a human, rather than AI, still matters to us.
As artificial intelligence finds its way into healthcare, therapy, and daily emotional support chatbots and other “AI tools,” questions about artificial empathy are at the heart of much debate. Can AI truly offer empathy? When it mimics empathy, do we receive, feel, and accept it in the same way as human empathy?
A new set of studies suggests that we feel more supported when we think the same response comes from a human versus AI.
The Three Main Types of Empathy
Our ability to receive empathy is our receptive affective capacity. There are three distinct types of empathy:
- Cognitive Empathy (“understanding of”): The ability to understand another person’s point of view or mental state (“perspective-taking”) helps communication and problem-solving.
- Emotional or Affective Empathy (“feeling with”): The capacity to feel what another person is feeling creates emotional resonance, fueling compassion and connection.
- Compassionate or Motivational Empathy (“caring about”): The desire to reduce distress motivates people to help others.
New Study Suggests Preference for Empathy Labeled as Human
A new set of studies published in Nature Human Behaviour investigates our preferences for empathic responses labeled as human- or AI-generated. Across nine large-scale studies involving over 6,000 participants, researchers explored how people perceive AI-generated empathic responses when they believe they are coming from a human vs. an AI model.
Participants valued empathy more when they thought the same answers came from a human, even though all answers were AI-generated. Participants received empathic responses from large language models (LLMs) like ChatGPT. The source of the message was labeled either as “human” or “AI.”
Despite all responses being AI-generated, those labeled as human were perceived as:
- More empathic
- More supportive
- More emotionally resonant
- More likely to make participants feel heard and less alone
- The human label made the most difference when responses involved affective (“feeling with”) and motivational (“caring about”) empathy.
This effect remained, even in short interactions, longer conversations, across different LLMs, and when there was an added three-minute pause before a response.
However, this does not mean that companies should simply label AI-generated responses as human to enhance reception. Outside of the research context, mislabeling or not informing users that responses are AI-generated would violate ethical responsibilities of transparency and informed consent.
But What If Human Responses Are Less Technically Empathic?
These findings are elucidating, especially when taken in conjunction with another study that found that people rate AI-generated responses as more compassionate than human-generated ones. Higher compassion scores for AI responses held even when people knew the response was AI-generated.
So what do we really want? Well, it likely depends on the person and situation.
We need more research to understand the conditions and factors that may determine our preferences to engage with human or AI-generated empathy.
For example:
- Does our attachment style impact whether we prefer AI or human empathy?
- What are the long-term impacts of relying on AI empathy in terms of loneliness and anxiety?
- Are people more comfortable discussing certain topics with AI instead of another person?
- Does the type or level of crisis or urgency, or the availability of alternative resources, influence this preference?
The Unique Value of Human Empathy
“Perfect” or frictionless empathy that is emulated by AI may not necessarily make us stronger in the long run. Our deep wish for human connection is innate, and the learning curve in accepting the limitations and “imperfections” of human empathy fosters deep personal growth and resilience over time.
Our relationships—and our ability to deal with complex social relationships—grow stronger from the unpredictable process that occurs between two people. Not only is there something uniquely healing about human empathy, but also navigating human relationships improves psychological flexibility and resilience.
Personal growth and resilience occur at the edges of “imperfect” empathy, where we begin to recognize and embrace the limitations and imperfections of what it means to be human, in others and ourselves. We let go of, rather than reinforce, a need for perfectionism or sole focus on efficiency.
There will still be important applications of AI-generated empathy, which offers advantages of 24/7 availability and broad knowledge (weighed against the risks of AI sycophancy, bias, and potential stigma). And AI tools have been shown to help people write more empathetically, so long as not all empathic writing is outsourced to AI. A recent study illustrates how relying on ChatGPT for writing decreases brain activation in certain areas and memory of what was written.
Knowing there is another person behind a message fosters a unique connection between two humans. This study reinforces the reality that the empathy and presence we can offer each other, albeit perhaps not as perfectly articulate as AI can be, is distinct and invaluable.
Copyright © 2025 Marlynn Wei, M.D., PLLC. All rights reserved.
