Empathy

Synthesis: Empathy enters education from three directions at once: as an outcome educators try to teach, as a construct they try to measure, and as something AI systems are increasingly said to display. The evidence is uneven across all three. Teaching empathy works — role-play, VR and LLM dialogue all produce measurable pre-post gains — but the effect shrinks as the experience becomes less embodied, and the comparison studies confound modality with time on task. Measuring empathy is harder than teaching it: observer ratings hold together, while computer-vision proxies agree on physical contact, disagree on eye contact, and invert on smiling. AI can rehearse empathic conversation at scale, but the best-controlled study shows self-assessed competence rising without an observed-performance check. And the newest concern runs the other way: believing AI can supply empathy makes people less willing to supply it themselves (Shastry & Gray, 2026).

Questions to Consider

• If role-play, VR and a chatbot all raise empathy scores, is the difference learning or time on task?
• A nursing student smiles at a patient in pain: observers read incongruity, a smile detector reads warmth. Which should an assessment trust?
• Most studies of AI-supported empathy training measure confidence rather than observed behavior. If confidence rises but behavior does not, has the training failed or merely not been tested?
• When a chatbot responds warmly, students report feeling understood. Is that empathy, a performance of empathy, or something that matters regardless of which?
• If people offload empathic responsibility once they believe AI can do it, what does that imply for a curriculum that teaches empathy partly through AI practice?
• Who should decide whether empathy is a gradeable outcome — and what happens to students who are accurate but not warm?

Introduction

Empathy is usually decomposed into at least two components: a cognitive component, perspective taking, and an affective component, sharing another's feeling, with some accounts adding a motivational component, the impulse to help. Personal distress — the aversive reaction to another's suffering — is often measured alongside and behaves differently from the rest. This matters for education because the components are teachable in different ways and measurable with different degrees of success.

AI enters at three distinct points, and conflating them is the main source of confusion in the literature.

1. AI as a teaching method for human empathy. Simulations, VR embodiment, role-play with LLM-generated personas, and generative imagery used to make another's experience vivid.
2. AI as a measurement instrument. Computer vision scoring body language in recorded simulations, or scales that treat empathy as a disposition expressed in AI interaction.
3. AI as an empathic actor. Chatbots that play patients, coaches or companions and are designed to seem understanding, which raises the question of whether the empathy is real, performed, or a substitute for human contact.

The first two are about learners; the third is about systems. A page that mixes them will produce claims like "AI improves empathy" that mean three different things.

Can empathy be taught?

The direct comparison is unusually clean in design and unusually confounded in execution. Zwoliński et al. (2026) randomly assigned 64 master's students in an accessibility course to real-world role-play with wheelchairs, aging suits and pregnancy simulators; the same equipment in a VR lab; or persona-building dialogue with ChatGPT-4.0. All three groups improved significantly on the Davis Interpersonal Reactivity Index. Role-play gained 13.2%, VR 11.6%, and the AI condition 8.8%. Only Personal Distress differed significantly between groups.

The ordering is intuitive and the interpretation is not. The three conditions ran for roughly 125, 87 and 42 minutes, so the study compares modality and dosage together. It is fair to say that a 42-minute LLM dialogue produced a smaller gain than a two-hour embodied exercise; it is not fair to say that embodiment is worth three times as much per minute. The authors are explicit about the confound, and their students' own recommendation — combine VR's repeatability, the real world's sensory richness and the LLM's adaptivity — is a reasonable reading of the evidence.

The health professions literature adds scale but not much causal clarity. Ismail and Zafira (2026) reviewed 17 studies of AI for affective-domain development and found the interventions concentrated in communication training, virtual patients, reflective feedback, breaking-bad-news simulation and OSCE-style assessment. Most reported gains in confidence, self-efficacy, reflective engagement or perceived communication skill. Evidence for empathy itself, emotional responsiveness, relational authenticity and non-verbal communication remained limited. That is the central finding: a field that intends to teach empathy is mostly measuring comfort.

A fourth strand makes the ambiguity explicit: wearable arousal sensors and AI-generated imagery can make another's circumstances vivid, but a rise in skin conductance is evidence of arousal, not of empathy, and studies in this line often stop there (Lim et al., 2024).

Can empathy be measured?

Observer ratings are the closest thing to a reference standard, and they behave reasonably. In Nissan et al. (2026), two faculty rated 37 nursing students in a recorded simulation with a 12-item inventory (α = 0.91) and a six-gesture scale (α = 0.93), and the gesture ratings correlated with the inventory at 0.41 to 0.63.

The algorithms told a more interesting story. Physical-contact detection agreed with four observer measures at 0.46 to 0.54. Eye contact and heatmap detection showed no significant association with observer judgments. And the smile detector correlated negatively with three observer measures, and with total empathy at r = −0.322, because the observers were watching a student smile at a patient in pain and reading it as nervousness or incongruity. The signal was real; the interpretation was backwards. This is the same inferential gap documented in sensor and gaze analytics, in a setting where the stakes are a clinical grade.

Self-report scales occupy the other end. The AI Empathy Scale for Generative AI Interaction treats empathy as a human disposition expressed during AI use and asks respondents about their own cognitive, affective and motivational responses; its preliminary factor structure did not reproduce the intended six dimensions (Kang et al., 2026). A disposition questionnaire and a behavioral observation are not interchangeable, and neither substitutes for patient ratings of whether the encounter felt caring.

Can AI rehearse empathy?

Yes, at least for the rehearsal part. Müller et al. (2026) gave 162 medical students one conversation with a role-prompted GPT-4o in one of four scenarios — shared decision-making, motivational interviewing, an STI consultation, breaking bad news — and self-assessed communication competence rose from 5.89 to 6.83 on a 0–10 scale (d = 0.58) after an average of 13 exchanges. Students rated the AI feedback useful at 7.92 out of 10, but feedback adequacy did not predict the gain (r = 0.08). There was no control group.

The pattern recurs: a short, low-stakes rehearsal produces a large self-efficacy gain, and the mechanism is more likely the rehearsal than the model's critique. That is not a reason to discard the tool. Standardized patients are expensive, scarce and scheduled; a chatbot is none of those, and a student who has already talked through a breaking-bad-news conversation once will use a standardized-patient session differently. It is a reason to stop calling the gain evidence of competence.

Where empathy is a conversational quality rather than a skill, the evidence is about perception. Wang et al. (2026) held model, platform and temperature constant and changed only the role and conversational moves in the prompt, producing a large difference in perceived empathy (21.27 versus 18.24, r = 0.53). Empathy, in the sense students experience it, can be scripted. Whether scripting it is educationally good is a separate question.

The substitution problem

The newest line of work asks what AI empathy does to human empathy. Across four studies with 249, 494, 743 and 737 participants, Shastry and Gray (2026) found that the more capable people believe AI is at providing emotional support, the less willing they are to provide it themselves — empathy offloading — controlling for trait empathy and prior AI use. When the recipient's source of support was named, participants said they would offload less to AI than to friends or therapists, because they rated AI's empathic capacity lower. On a behavioral measure, the same participants who said they would not offload chose fewer empathy words and more indifference words.

For education this cuts two ways. A curriculum that uses AI to practice empathy may be building the skill; a campus culture that routes emotional support to chatbots may be eroding the motivation to use it. The two effects are compatible, and nothing in the current literature measures them together.

What this means for practice

• Separate the three questions. When someone claims AI improves empathy, ask whether they mean teaching humans, measuring humans, or simulating empathy. The evidence differs for each.
• Do not accept self-efficacy as an empathy outcome. Confidence is cheap and rises after almost any rehearsal; observation, standardized patients or patient ratings are the measures that matter.
• Treat vision scores as second readers. Physical-contact detection is usable for formative feedback; smile and eye-contact detection are not, without scenario context a model does not have.
• Budget time, not just modality. The comparison studies confound the two, and a cheaper method that takes a quarter of the time may still be the better choice — just not an equivalent one.
• Watch for offloading. If AI becomes the default channel for emotional support, the empathy you are trying to teach may be quietly outsourced.

Common Misconceptions

• "AI can teach empathy." AI can support rehearsal and Reflection; the strongest review finds limited evidence for empathy as an outcome, and the best-controlled training study measures self-assessment.
• "Empathy is warm language." Perceived empathy can be produced by prompt design alone, which is evidence about perception, not about understanding.
• "A smile means warmth." In a distressing clinical encounter, observers read smiling as incongruity, and a facial-action detector reads it as positive affect. Context is the missing variable.
• "Self-report scales measure empathy." They measure a disposition or a self-concept; they do not measure behavior in an encounter.
• "More immersive is always better." Immersive methods show larger gains, but at several times the cost and duration, and the comparisons do not equalize exposure.

Connected Concepts

• Affective Tutoring
• Affective Computing
• Social-Emotional Learning
• Workplace Learning
• Simulation
• Medical and Health Professions Education
• Nursing Education
• Sensor and Gaze Analytics
• Assessment Validity
• Self-Report Measures
• Human AI Collaboration
• AI Sycophancy
• Pedagogical Safety
• Well-Being
• Arts, Design and Media Education

Connected Articles

• Teaching Empathy: A Comparative Analysis of Real, Virtual, and AI-Driven Methods — Role-play 13.2%, VR 11.6%, AI dialogue 8.8% on the IRI, with exposure time confounded (Zwoliński et al., 2026)
• Detecting Nursing Students' Empathy in Video-Recorded Simulation Using Computer Vision — Physical-contact detection agreed with observers; smile detection ran negative (Nissan et al., 2026)
• Training Empathetic Communication Skills in Medical Students With a Role-Prompted GPT-4o Chatbot — Self-assessed competence rose d = 0.58 after 13 exchanges, with no control group (Müller et al., 2026)
• Artificial Intelligence for Affective-Domain Development in Healthcare Professions Education: A Systematic Review — 17 studies; confidence gains common, empathy evidence limited (Ismail & Zafira, 2026)
• Exploring the Potential of AI in Nurturing Learner Empathy, Pro-Social Values and Environmental Stewardship — EDA and generative imagery as an arousal pipeline, with empathy claimed rather than measured (Lim et al., 2024)
• People offload empathy when they think AI can give it instead — Belief in AI's empathic capacity predicted less willingness to provide support (Shastry & Gray, 2026)
• AI Empathy as a Human Disposition in Generative AI Interaction: Development and Preliminary Psychometric Evaluation of a Cognitive–Affective–Motivational Scale — A 42-item disposition scale whose factor structure did not match its design (Kang et al., 2026)
• Invisible Impact of Empathy on Behavioral Change: Isolating the Effect of Empathy in Long-term Physical Activity Coaching Chatbot Interactions — Empathetic language raised felt understanding, which predicted engagement rather than behavior change (Siyan et al., 2026)
• Comparing teacher-centered and student-centered agents based on prompt engineering: Effects on learning performance, cognitive load, flow experience, and empathy perception in physics learning — Prompt-specified role and moves produced a large perceived-empathy difference (Wang et al., 2026)
• Nursing Students' and Faculty Experiences with Artificial Intelligence in Education: A Qualitative Study Using the Technology Acceptance Model — Nursing students described AI as emotional support during clinical stress (Akbaba et al., 2026)