Source: Rubin et al. (2025) — “Comparing the Value of Perceived Human versus AI-Generated Empathy,” Nature Human Behaviour
AI can now write very comforting sentences.
Sometimes too comforting. The kind of sentence that makes you think, “This is emotionally intelligent,” followed immediately by, “Wait, does it actually care, or is it just very good at sounding like it cares?”
This paper studies perceived human versus AI-generated empathy. Across studies, empathic responses were labelled as coming from either humans or AI, and people evaluated them differently.
The core tension is fascinating.
AI can produce responses that look empathic. It can acknowledge feelings, reflect pain, offer support, and avoid saying the terrible things humans sometimes say when trying to help. Honestly, some humans could use the training data.
But people still care about attribution.
A response believed to come from a human can feel more valuable because empathy is not only about the words. It is also about the belief that another conscious person noticed you, cared, and chose to respond.
That is hard for AI to fully replicate.
This matters for emotional support, healthcare, customer service, coaching, and any context where organisations are tempted to scale empathy. Scaling empathy sounds wonderful until you realise empathy may lose something when people know it is automated.
The practical question is not whether AI can write supportive text.
It can.
The question is what kind of support people think they are receiving, and whether disclosure changes the meaning of the interaction.
My takeaway: AI empathy may be useful, but it is not the same social object as human empathy. The words matter, but so does the imagined listener behind them. Sometimes people do not just need a good response. They need to believe someone is there.