968 people were asked to comfort someone going through a hard time. 74% said they encouraged the other person to open up. Independent evaluators found only 18% actually did.
The data comes from a preregistered experiment at Northwestern and Stanford, conducted on a platform called “Lend an Ear”. Participants offered empathic support across 2,904 text-based conversations with AI partners simulating real personal and workplace troubles: a family member diagnosed with cancer, losing a job, being passed over for a promotion. 91% of participants rated the scenarios as “quite a bit” or “very much” realistic.
The core finding: there is almost no relationship between feeling empathic and communicating empathy. The researchers measured trait empathy using two validated psychological scales and then scored participants’ actual conversations across six dimensions of empathic communication. The correlation was near zero (R-squared = 0.001). People who scored high on standard empathy scales performed no better in conversation than those who scored low.
The gap has a name
The researchers call it the “silent empathy effect.” You feel it. You just don’t transmit it.
Most participants were confident they did well. The gap between self-assessment and measured performance was not random. It was systematic. People consistently overestimated how much they encouraged others to share, how much they validated emotions, and how well they demonstrated understanding.
For a CEO managing 80 to 200 people, the finding should be uncomfortable. Your managers likely care about their teams. They probably score well on any empathy assessment HR runs. But the research suggests caring and communicating care are different skills, and most people have only developed one of them.
What the AI coach actually did
In a randomized experiment, participants who received personalized feedback from an AI coach improved by nearly a full standard deviation (0.98 SD) after a single practice session. 21.6% showed reliable improvement, compared to 2.9% in the control group. A video-only training group showed marginal improvement (9.0%), confirming that the personalized, real-time feedback was the active ingredient, not just exposure to empathy concepts.
The AI coach didn’t teach participants to be more empathic. It taught them specific conversational moves that make their existing empathy visible.
Three prescriptive behaviors improved significantly: encouraging elaboration (asking the other person to share more), validating emotions (acknowledging what the other person feels), and demonstrating understanding (paraphrasing to show comprehension). Three counterproductive behaviors decreased: giving unsolicited advice, shifting focus to oneself, and dismissing the other person’s emotions.
Independent raters confirmed the shift. When shown pairs of conversations, one pre-training, one post-training, without knowing which was which, they preferred the post-training conversation roughly two-thirds of the time.
128 ways to make someone feel heard
The researchers didn’t stop at measuring the gap. They mapped 128 distinct communicative patterns from 16,975 human messages using machine learning on text embeddings. The patterns organize into a four-level taxonomy: affective empathy (sharing emotions, 25% of messages), cognitive empathy (recognizing emotional states, 27%), motivational empathy (expressing care and willingness to help, 26%), and misattuned behaviors (22%), the responses that feel helpful to the person giving them but don’t land with the person receiving them.
The most common misattuned behavior was advice-giving. Participants meant well. They wanted to fix the problem. But telling someone “you just need to move on” when they’ve just lost a family member doesn’t make them feel heard. It makes them feel managed.
After AI coaching, advice-giving dropped by 3.8 to 5.1 percentage points. Validating emotions increased by 2.9 to 3.9 percentage points. The quality of what participants said changed. The quantity didn’t. Conversations were the same length before and after training. People weren’t talking more. They were talking differently.
What this means for organizations
The paper reframes empathy from a personality trait into a set of learnable communication skills. The reframing matters for anyone who manages people.
Most leadership development programs treat empathy as a disposition. You either have it or you develop it over years of practice and reflection. The research points to a different model: the disposition is already there in most people. What’s missing is the translation layer, the specific moves that turn internal empathy into something the other person can actually feel.
One AI coaching session. Not a six-week workshop. Not a certification program. One session produced measurable, independently verified improvement in how people communicate care.
The question for any leader managing a team: if your managers already feel empathy but their teams can’t tell, is that a character problem or a skills gap? And if it’s a skills gap, how long are you willing to leave it untrained?
Related: 66% of CEOs Are Freezing Hiring While Betting Billions on AI explores another gap between intention and execution: the disconnect between AI investment and organizational redesign.