AI can predict how you feel better than most humans, but doesn’t actually feel anything. Does that matter? How? Studies show it outperforms crisis workers at validation – and you can tune it.
The story
What I learned about AI and “empathy” kind of broke my brain [in a good way].
When we say empathy, we usually mash three things together: – Cognitive empathy: I can *accurately guess* how you feel. – Affective (emotional) empathy: I actually *feel* some of what you feel. – Motivational empathy: I feel moved to *help* you.
AI only has the first one. Cognitive. And it’s weirdly good at it.
In several studies, large language models [think ChatGPT] wrote replies to people in distress that third-party humans rated as *more* understanding, validating, and caring than real crisis line workers. Especially for painful stuff.
Why? The model: – Predicts your inner state from patterns in your words/voice – Paraphrases your experience clearly – Reflects your feelings back consistently [no “sorry, bad day at the office” variability]
But it’s not *feeling* anything. No shared gut-punch. No inner pull to help. Just scary-good pattern recognition.
Here’s the fun part: you can aim that cognitive empathy.
Instead of “Be nice,” try chunked prompts like: – “First: infer the user’s likely emotions and needs from their text.” – “Second: validate those emotions in simple language.” – “Third: offer one or two gentle next steps.”
I’ve been building multi-agent setups [like my AgentFlow framework] where: – One agent decodes emotion/state – Another crafts the response – A third tunes the “vibe” for warmth and clarity
Is it perfect? No. It still struggles with unspoken cues and complex social history.
But if you treat empathy as a skillful workflow, not a magical trait, AI bec
Recent Comments