More people are turning to artificial intelligence (AI)-powered chatbots for companionship and to address loneliness. But can AI-generated empathy replace human connection?
In a study published in the journal Communications Psychology, a team led by researchers at Penn State found that human users rated AI-generated empathetic responses higher in quality than human-generated empathetic responses, yet they preferred human empathy over AI. In the following Q&A, C. Daryl Cameron, the Sherwin Early Career Professor and associate professor of psychology at Penn State, and doctoral student Joshua Wenger discussed AI empathy and why it may prove an effective complement to — and not a replacement for — human companionship.
Q: What does your work reveal about how people choose to use AI?
Cameron: In our work, we took a research approach that allowed people to show through their choices what they actually want and care about. Studying people's actual choices in the moment can sometimes reveal their preferences. That was a core part of our study. What's interesting is how, on the one hand, people don't seem to want AI empathy, but there is some subset of the population that does want it, so it's complicated. Nevertheless, people think that the actual content of what the large language models are saying is empathetic, interesting and validating, and so there's this paradox between what people want and what they choose to opt into, versus their actual evaluations of the material.
Wenger: We called this effect the “AI empathy choice paradox,” and it’s this consistent pattern where people rate AI empathy or empathetic expressions from AI as higher in quality and more empathetic than those from real humans. It makes people feel more heard, cared for and validated than responses from human empathizers. But despite this higher quality, people still prefer to receive empathy — even lower quality empathy — from humans over AI.
Q: What drives this preference for human or AI empathy?
Wenger: For every response in our study, we asked people to rate how empathetic they found each response. We also asked them how much effort they thought the empathizer put into writing the response. This builds upon some theoretical work by the scholar Anat Perry and her paper “AI will never convey the essence of human empathy.” In it she argues that one of the more important aspects of human empathy is the exertion of effort. Daryl’s work has shown that empathy is an effortful act. The idea is that I, as a human empathizer, expend my emotional resources to express care for you, and by doing so, I am signaling your importance to me in some way. And because empathy is objectively effortless for AI, it's not emotionally effortful.
Interestingly, what our paper found is that when we asked people to rate the perceived effort in the AI- and human-generated responses, they rated AI overall as exerting more effort in the creation of these responses. What this means is complicated, because it's unclear whether people were answering the question as if AI were a human or genuinely perceiving effort. Regardless, we found this difference where, for AI empathy, it seemed like the empathizer put a lot of work into their response. In addition, we found that the more effortful a response is perceived as being, the more empathetic people tend to perceive it. So, if someone puts a lot of effort into writing a response, it tends to make the response seem more empathetic. That could be one driver of the high-quality AI empathy.
Q: Do people who regularly use AI tend to prefer AI empathy over human empathy?
Cameron: You might think that people who use AI more frequently would be more willing to choose AI empathy, but we don't tend to see that. Instead, we found these interesting patterns where AI empathy is rated as exerting more effort, but the relationship between effort and empathy also depends upon who's the one doing it. The relationship is stronger for humans than it is for AI. As you see higher levels of effort exerted, human empathizers catch up to AI. So, the gap between human and AI empathy is much smaller at higher levels of effort.
There's a broader, big picture story here: How can we learn from our interactions with AI? For instance, if I’m tired and feel like I cannot write another condolence message, perhaps I can use AI to guide me.
Wenger: In one of our studies, we compared AI to responses that had been previously generated from trained crisis responders. Even these empathy experts still weren't catching up to AI in empathic quality most of the time. Trained crisis responders are expert empathizers, but there's still no relational context to the individual seeking empathy. It could be different when you talk about empathy from your romantic partner compared to AI.
That being said, we’re seeing companionship increasingly as a common use of AI. People are forming friendships and romantic relationships with AI. It brings up questions such as, if it's an AI that you're interacting with over time, an AI therapist or an AI friend, how does that impact the quality of the empathy?
Q: What does the future of empathy look like in the AI era?
Cameron: One of the tradeoffs in AI empathy debates is the idea that the empathy most of us crave comes from the deep, rich conversations we have with people we know and who would know us individually, the people who choose to be with us. These individuals have a partial connection to us. How do people prioritize that partial tie, that connection to them individually, as unique persons? It may be that I'm willing to sacrifice some degree of empathetic quality if it's coming from someone I know really well, even if I know my best friend is not the greatest empathizer.
Would you prioritize that connection versus moments when you really need in-the-moment empathetic support? It's an open question. Assessing preferences carefully is a key part for understanding this debate, and so much of the focus has been that it's AI, it's not embodied, it's not real, it can't feel and therefore anything that looks like empathy doesn't matter as much. But I think we've tried to argue that it's important to take people's lived experiences at their word, too. So, if someone says they think that AI empathy expressions are valuable to them, I think we can try to be inclusive and say that those expressions could be possible for that person, and by assessing their motivations and their choices and their ratings, we're trying to give credence to that.
We're trying to explore the recipient side of empathy as an equal part of this whole exchange between the giver and the receiver. Let's focus on the receiver, even if they know the sender isn't human. They still see some value in the response. There are some interesting theoretical advances and methodological psychological science advances we can make that push at the boundaries of what we think empathy could be.
Wenger: Some of the pushback towards AI empathy, or empathy from an AI source, might be built upon this fear of replacing human empathy. In our studies, we framed it as a direct choice between a human and an AI empathizer. But it's also important to consider how these choices take shape in the real world. Maybe some of the greatest value that AI empathy could hold is when there's not actually a choice between a human and an AI empathizer. It could be that you turn to AI for empathy when there's no human available as an alternative source. So, considering the availability of empathizers, AI is not replacing human empathy but potentially supplementing it when human empathy is not present.