Practicing with Language Models Cultivates Human Empathic Communication
Aakriti Kumar, Nalin Poungpeth, Diyi Yang, Bruce Lambert, Matthew Groh
cs.CL, cs.HC
2026-03-16
They built a platform where humans support GPT-4o role-players and get LLM feedback; in a 968-person RCT the AI coach lifted empathy ~1 SD without flattening individual voice.
There is a counterintuitive result in the literature: when human-written comfort and LLM-written comfort are judged blind, the LLM often scores as more empathic. But the moment a recipient learns a message came from AI, the feeling of being heard drops. Outsourcing empathy to a machine makes receiving care feel worse.
So this paper is not about making AI comfort sound more moving. It asks the inverse: can AI teach humans to be more empathic themselves? Empathy is a learnable skill with large individual variation, which makes a coaching setup plausible to begin with. A team from Northwestern's Kellogg School and Stanford's CS department built a platform to test it in a randomized trial.
The platform, Lend an Ear, is text role-play: a human plays the supporter, and GPT-4o plays someone in trouble. The troubles come in five flavors, three workplace (job loss, passed over for promotion, feeling undervalued) and two personal (a family member diagnosed with cancer, a family member passing away). Each conversation runs about four minutes across roughly eleven turns.
The actual intervention is a second GPT-4o, a communication coach that gives personalized feedback after each conversation against a pre-registered six-dimension rubric: three prescriptive (encourage the other person to elaborate, validate their emotions, paraphrase to show understanding) and three proscriptive (avoid unsolicited advice, avoid steering the topic back to yourself, avoid dismissing the emotion). These six are not invented for the paper; they map onto empathy behaviors validated in clinical communication research.
The experiment is a 968-person RCT, recruited through Prolific and demographically representative of US adults, split into four arms: no-feedback control, a 92-second instructional video, AI-coach feedback, and video plus coach. Each participant completed three conversations, yielding 2,904 conversations and 33,938 messages (16,975 from humans). Every participant response was scored by an LLM-as-judge across the six dimensions.
Going deeper, they ran k-sparse autoencoders (kSAE) over 29,520 sentence units to mine features, surfacing 128 latent concepts nested into a four-level taxonomy: affective empathy (25% of messages), cognitive empathy (27%), motivational empathy (26%), and misattuned behaviors (22%). The taxonomy is discovered bottom-up from data rather than hand-listed ahead of time.
Effects are gated by the reliable change index (RCI), which counts a change as real only if it is too large to be measurement noise:
| Arm | Reliably improved | Empathy score gain |
| No-feedback control | 4.5% | baseline |
| Instructional video (92 s) | 9.0% | +0.55 SD |
| AI coach | 21.6% | +0.98 SD (2.9 pts) |
| Video + AI coach | 26.3% | +1.26 SD |
The coach produced significant gains across all six dimensions; the video moved only four. The biggest swings were behavioral: among the prescriptive ones, encourage elaboration rose 0.59 SD, validate emotions 0.47, paraphrase understanding 0.46; among the proscriptive ones, give advice fell 0.57 and dismiss emotions fell 0.43.
The self-knowledge gap is the sharpest finding. Participants' self-rated empathy correlated almost zero with actual performance (R² between 0.000 and 0.004). 74% said they encouraged elaboration quite a bit or more, and 87% said they showed understanding, yet the LLM judge rated only 18% and 9% as actually doing so. People think they comfort well; the numbers disagree.
The intervention did not flatten anyone. Semantic novelty held steady (medians 0.402 to 0.435 across arms), with only minor convergence in the video and combined arms on the third conversation. The spread between participants (0.178 to 0.203) was far wider than the spread between AI models, meaning humans kept sounding like themselves. An AI detector classified 90% of human supporter text as human-written, and the share falsely flagged as AI was lower in the coach arm (6%) than in control (14%). Coached answers did not start sounding robotic.
The scoring itself was cross-validated: 150 people made pairwise comparisons, and the Bradley-Terry ranking of human preferences correlated with the LLM's Elo scores at Spearman ρ=0.85.
This paper sets a reusable template for AI coaching humans on soft skills: the model gives feedback, but a human still does the talking. It sidesteps the opening paradox directly, since AI speaking for you makes the recipient feel worse, so the model steps back into a coaching seat instead of acting as a stand-in.
The six-dimension pre-registered rubric plus LLM-as-judge is portable to almost any setting where you measure and train human communication: negotiation, interviewing, performance reviews, customer-support coaching. The sharpest lesson for managers is simpler: do not trust self-assessments like saying you are a good listener. Participants wildly overrated their empathy, and self-rating tracked real performance almost not at all. To talk about communication quality you need an external standard, not a self-report.
The whole intervention is cheap; a few-minute session already moves the needle. But these are single-session, immediate effects, not a shipped product.
The setting is deliberately narrow. The experiment covers only low-familiarity relationships: strangers, acquaintances, workplace colleagues. The authors are explicit that they did not touch the thick empathy of family, partners, or close friends, which is where empathy matters most. The same playbook may not transfer cleanly to intimate relationships.
Breadth is limited on two more axes. The sample is US adults only, and empathy expression is heavily culture-bound. Only five trouble scenarios were tested, so the generalization boundary is unclear. The intervention ran once, with no follow-up; whether the gains stick or fade is unanswered here.
The scoring method deserves the most scrutiny. Every headline effect was measured by an LLM-as-judge. The authors validated it against human preference (Bradley-Terry ρ=0.85), but the effect sizes are still AI-graded. The conclusion that people think they are empathic but are not leans entirely on the judge ruling that participants failed, and if the six-dimension rubric has blind spots, part of that gap is the rubric's bias, not the human's.
And the old question: are coached responses genuinely more empathic, or just better at hitting rubric boxes? The authors argue on a spectrum with structured clinical training, but they did not test authenticity here.