AI chatbots can match or outperform people at emotional support, study finds

Across five experiments involving 1,233 participants, chatbot responses were rated more supportive for anger and fear, while AI and human messages performed equally well for sadness.

The study compared emotional support written by people with responses generated by ChatGPT, Claude and Gemini

People facing everyday feelings of anger or fear rated emotional support written by AI chatbots more highly than messages from human participants, according to research led by The University of Manchester in collaboration with Durham University.

The advantage did not extend to every emotion. AI-generated and human-written messages were equally effective in scenarios involving sadness, while the researchers stress that the study examined everyday, non-clinical support rather than professional mental health care.

What made the clearest difference was not whether the message came from a machine or a person. Specific, actionable advice, including breathing exercises, ways to reframe a thought and practical next steps, improved emotional outcomes across both sources.

The research comprised five experimental studies involving 1,233 participants. It compared responses from ChatGPT 4o, Claude 3.5 Sonnet and Gemini 1.5 Pro with messages written by people responding to emotionally charged situations.

Scenarios included being passed over for a promotion, returning home to find a shared house left untidy and learning that an important presentation would have a larger audience than expected. These represented sadness, anger and fear respectively.

Participants were not told who or what had produced each response. Many correctly identified whether a message came from an AI or a person, but recognizing the source did not reduce its benefit in the studies where large language model responses received higher support ratings.

Practical advice, not validation, drives stronger results

The large language models broadly replicated the strategies people use to manage someone else’s emotions. Of the three models tested, ChatGPT displayed the greatest similarity to human response patterns.

However, simply acknowledging a person’s feelings did not account for the models’ stronger performance. Adding or removing emotional validation made no difference to how helpful participants found the messages.

Actionable guidance proved more consequential. Responses became more effective when they gave people clear steps they could use, regardless of whether the writer was human or AI.

Dr Belén López-Pérez, Lecturer in Psychology at The University of Manchester and a co-author of the study, says: “Our findings suggest that what makes emotional support effective is less about who provides it, and more about how it is delivered. Providing people with specific, actionable steps, rather than general reassurance, is a key ingredient for effective support, whether that comes from a person or a chatbot.”

She adds: “While a human might typically suggest going for a walk or making a cup of tea, an AI chatbot tends to go further, for example acknowledging the person’s feelings and offering step-by-step strategies such as breathing techniques or specific ways to reframe a thought. Training human supporters – whether professionals or members of the public – to provide this kind of structured, actionable guidance could meaningfully improve the quality of everyday emotional support.”

AI support does not replace human presence

The findings show that chatbot responses can sometimes improve a person’s emotional state more effectively than messages written by another person, but the result depends on both the emotion involved and the outcome being measured.

Yuhui Chen, a PhD researcher at The University of Manchester and co-author of the study, says: “We found that AI can emulate the supportive strategies that humans use, and at times even outperform human supporters in improving a person’s emotional state – though this is not universal. The effect depends on the type of emotion and what outcome you are looking at.”

She adds: “We are not saying that humans should or can be replaced by AI. But our findings do show that when people choose to use AI for emotional support, it can be genuinely helpful, and understanding why it works can help us improve how both humans and AI provide that support.”

Dr Sarah Walker, a co-author from Durham University’s School of Education, identifies a limitation that message ratings alone cannot capture. An AI system can recommend an action, but it cannot provide the time and physical presence that may accompany support from another person.

“When AI suggests a walk or a breathing exercise, the suggestion is where its involvement ends. When a person offers the same thing, it usually comes with their presence as well. We don't just suggest the walk, we go on it,” she says.

“That kind of support takes real time and effort, and the people closest to us are often stretched thin. So to me, this work is less about AI replacing us and more of a reminder that what we want from each other in hard moments is often more than the people we love can give.”

The peer-reviewed paper, “A Digital Shoulder to Cry On: Understanding Why Large Language Models Can Be Effective in Extrinsic Interpersonal Emotion Regulation,” is published in the journal Emotion.