Gen AI for Emotional Support: Do AI replies match or exceed human empathy in mental health scenarios?

Overview

A new study from the Universities of Manchester and Durham examines how generative AI messages compare with human replies when offering emotional support. Across five imagined scenarios, AI responses were rated as more emotionally supportive for anger and fear, while sadness showed no reliable difference. Researchers highlight actionable, concrete suggestions as a key driver of comfort, regardless of the source, and suggest that AI can help with word choice while humans retain advantages in ongoing interactions. The study also notes limits, including imagined scenarios and brief exchanges, and calls for caution about long-term dependence or crisis contexts. Author: The Conversation.

Key insights

• AI messages were seen as more empathic in anger and fear contexts; sadness showed no clear difference.
• Concrete, actionable advice boosted comfort regardless of source.
• Perception matters: describing a message as human-written increases perceived empathy compared with AI attributions.
• Humans remain preferred for deep, ongoing emotional engagement despite AI advantages.

Overview and context

The Conversation team reports on five experiments evaluating whether generative AI (gen AI) writing emotional-support messages can rival or exceed human-written responses. The work, a collaboration between the universities of Manchester and Durham, builds on growing use of gen AI for mental health support, including situations where individuals seek help after a rotten day or when anxious about a presentation. The study finds that gen AI messages can be rated as more emotionally supportive than human messages in certain scenarios, but not uniformly across all emotions.

Methods and scenarios

In one core experiment, 390 participants imagined situations involving anger, sadness or fear and read either human-written or gen AI responses without being told the source. The design allowed a direct comparison of perceived support without source bias. Additional experiments explored whether the AI advantage persisted when message content emphasized empathy, validation of feelings, or practical, actionable steps. The researchers also examined how the recipient’s perception of the helper’s effort and relationship context influenced outcomes.

Key findings

Across the anger and fear scenarios, gen AI messages tended to be rated as more emotionally supportive than human replies, while for sadness there was no reliable difference. Gen AI messages sometimes increased calm in fear contexts more than human messages, but did not consistently reduce fear itself. A broader pattern emerged: messages that offered concrete, realistic actions were judged as more comforting and improved some emotional responses, regardless of whether the sender was a gen AI or a person. When human messages provided similar actionable support, they were judged as similarly comforting to AI messages in those comparisons.

The study notes that the wider literature often finds gen AI messages to be perceived as empathic, sometimes more so than human responses, but this advantage can wane when recipients know the message came from AI. Across nine further studies, gen AI replies were rated as more empathic when described as human-written rather than AI attributions, and people tended to prefer human interaction for emotional engagement, even when it means waiting longer.

The practical lesson and limits

The researchers stress a practical takeaway: when someone is upset, ask if they want advice, and if so offer a single, manageable next step rather than a long list of remedies. Gen AI can help by suggesting useful wording and potential responses, but human relationships bring personal knowledge and ongoing involvement that AI cannot replicate. The authors acknowledge limits: many experiments used imagined, brief exchanges and could not simulate the long-term use of chatbots or crisis scenarios. The messages used in the studies were often provided by strangers or crafted as representative examples, rather than originating from a friend who truly knows the recipient.

Implications for practice and policy

Used carefully, gen AI could complement human support by providing structured, empathic language and proposing actionable steps. The findings suggest the value of combining AI-enabled drafting with human empathy and ongoing presence, rather than replacing human support entirely. The study also highlights the importance of source attribution and the recipient’s perception of the helper’s time and emotional investment in shaping responses and relationship satisfaction.

Limitations and future directions

The wider research base remains limited in long-term and crisis contexts. Future work should explore how sustained use of gen AI influences mental health outcomes and how AI can best integrate with real-life friendships and clinical support networks. The study closes with a cautious note: gen AI can be a useful tool for improving conversational skills and expanding the repertoire of supportive words, but it cannot substitute the personal knowledge and continued involvement of someone close.