What is it about?

Giving AI a voice can make people feel more heard, but it can also backfire. Across five studies, we examined how vocal expression impacts perceived empathy for humans and AI. Critically, the interaction context mattered: 1) In asynchronous interactions with prerecorded messages, voice reduced perceived empathy and feelings of being heard compared with text. This happened for both human and AI responses, but the penalty was especially strong for AI. This appeared to be because AI voices evoked feelings of uncanniness, hindering effective empathy. 2) In live, back-and-forth conversations with AI about participants’ own experiences, voice no longer reduced empathy or feeling heard—and instead sometimes even improved both. Importantly, in these live conversation contexts, voice was not seen as more uncanny. These results foreground the possible double-edged and contextual nature of immersive design features.

Featured Image

Why is it important?

People are turning to AI for emotional support more than ever. A lot of past work highlights the impressive quality of AI empathy compared to human empathy; however, little work has engaged with how new, cutting-edge features of these emerging technologies shape our empathic experience. Specifically, AI voices have become more powerful and realistic than ever, leading us to wonder what impact this might have on people seeking emotional support. Additionally, for those designing AI systems that offer emotional support, these findings suggest that adding voice deserves careful testing in the context where people will actually use it.

Perspectives

I believe this speaks to the nuanced and contextual nature of empathic AI. There are many competing dynamics in AI empathy. For example, people generally rate AI responses as higher quality, yet still choose to avoid them in favor of human empathy. Likewise, immersive design features such as voice can elicit different reactions depending on the context: they can feel aversive and uncanny in asynchronous communication, yet facilitate effective empathic communication in multi-turn, free-form dialogue. This highlights the importance of considering how context shapes our communication with and emotional responses to these AI systems.

Joshua D. Wenger
Pennsylvania State University

This paper is exciting to me because it reveals the complexities of our choices about empathic AI. In order to make informed scientific and moral judgments about empathic AI, we should try to understand that different people want varying things from these platforms. Just as there is variety of motives for engaging, there is variety in the types of AI environments people have before them. Discussions of AI empathy should appreciate this diversity of what it can be like to receive empathy from artificial intelligence.

Daryl Cameron
Pennsylvania State University

As a Human-Centred AI researcher with an engineering background, this work made me think again about how important well-thought-out interaction design is, beyond just the technical capabilities of a system. One implication of the findings of this paper that comes to mind is how we interact with characters in video games. An NPC might have a realistic, expressive voice, but that alone doesn't make it feel caring or believable. Imagine failing a mission and hearing, "I'm sorry that happened. There is always a next time." If it's just a recorded line and the character can't respond to anything you say back, the interaction might feel awkward or even break your immersion. Now imagine being able to talk with that NPC about what went wrong and having it respond to what you actually say. The voice could be exactly the same, but the quality of experience and perceived empathy would feel very different.

Reza H. Mogavi
University of Toronto

Read the Original

This page is a summary of: Voice can help and hurt AI empathy: How context shapes the effects of vocal delivery., Technology Mind and Behavior, October 2026, American Psychological Association (APA),
DOI: 10.1037/tmb0000235.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page