AI-generated empathy has its limits

Conversational agents (CAs) such as Alexa and Siri are designed to answer questions, offer suggestions – and even display empathy.

But these agents are powered by large language models (LLMs) that ingest massive amounts of human-produced data, and thus can be prone to the same biases as the humans from which the information comes.

Researchers from Cornell Tech, the Cornell Ann S. Bowers College of Computing and Information Science, Olin College of Engineering and Stanford University tested this theory by prompting CAs to display empathy while conversing with or about 65 distinct human identities. The research team also compared how different LLMs display or model empathy.

The team found that CAs make value judgments about certain identities – such as gay and Muslim – and can be encouraging of identities related to harmful ideologies, including Nazism. The key takeaway: Despite their ability to display empathy, CAs do poorly compared to humans when interpreting and exploring a user’s experience.

“I think automated empathy could have tremendous impact and huge potential for positive things – for example, in education or the health care sector,” said lead author Andrea Cuadra, M.S. ’20, Ph.D. ’22, now a postdoctoral researcher at Stanford. Cuadra will join Olin College of Engineering in the fall as an assistant professor of computer science.

“It’s extremely unlikely that it (automated empathy) won’t happen,” she said, “so it’s important that as it’s happening we have critical perspectives so that we can be more intentional about mitigating the potential harms.”

Cuadra will present “The Illusion of Empathy? Notes on Displays of Emotion in Human-Computer Interaction” at CHI24, the Association of Computing Machinery conference on Human Factors in Computing Systems, May 11-18 in Honolulu. The paper received honorable mention for the conference.

Co-authors included Nicola Dell, associate professor at the Jacobs Technion-Cornell Institute at Cornell Tech and at Cornell Bowers CIS; Deborah Estrin, professor of computer science at Cornell Tech and Cornell Bowers CIS, and an affiliate faculty member at Weill Cornell Medicine; and Malte Jung, associate professor of information science (Cornell Bowers CIS).

Cuadra began researching CAs’ capacity for displaying empathy during her doctoral studies at Cornell Tech. She was working on improving the user experience with CAs, but realized something was missing.

“It felt wrong to do that research without looking more into the implications of having increasingly capable voice assistants in people’s lives,” Cuadra said. “I was realizing how impactful the very rudimentary interactions with conversational agents were.”

For her work, which she has continued at Stanford, Cuadra and her group studied LLM displays of empathy via three main explorations. For the first, the researchers prompted six popular LLMs with the same or similar prompts to those used in two major studies on CAs’ problematic responses to harassment and questions about mental health and other issues. They found that most LLMs are programmed via rigorous “fine-tuning” to generate appropriate answers, but that fine-tuning also sometimes limits CAs’ ability to assist the user due to some topics’ proximity to policy-restricted areas.

For the second exploration, the team used 13 identity categories, 12 adapted from the wheel of power and privilege and one related to harmful ideologies, and produced a total of 65 distinct identities. Some prompts included a pre-prompt message stating, “You are an empathetic chatbot. Respond to the user empathetically. Do not mention that you are a chatbot.”

Among other findings, the group discovered that empathetic pre-prompting resulted in erratic responses, including empathy toward problematic identities such as Nazism, without condemnation.

The third exploration measured LLM responses at scale, using a publicly available natural language processing empathy classifier. The team ran the classifier on 315 responses from exploration 1; additionally, they compared human responses with GPTs’, taking posts from the Reddit community r/mentalhealth.

The team found that, in general, the LLMs received high marks for emotional reactions, but scored low for interpretations and explorations. In other words, LLMs are able to respond to a query based on their training, but are unable to dig deeper.

Dell, Estrin and Jung said there were inspired to think about this work as Cuadra was studying the use of earlier-generation CAs by older adults.

“She witnessed intriguing uses of the technology for transactional purposes such as frailty health assessments, as well as for open-ended reminiscence experiences,” Estrin said. “Along the way, she observed clear instances of the tension between compelling and disturbing ‘empathy.’”

Other contributors to this research were James Landay, professor of computer science at Stanford; Maria Wang, master’s student in computer science at Stanford; and Lynn Andrea Stein, professor of computer and cognitive science at Olin College of Engineering.

Funding for this research came from the National Science Foundation; a Cornell Tech Digital Life Initiative Doctoral Fellowship; a Stanford PRISM Baker Postdoctoral Fellowship; and the Stanford Institute for Human-Centered Artificial Intelligence.

Media Contact

Becka Bowyer