People offload empathy when they think AI can give it instead

Synthesis: Shastry and Gray (2026) ask what happens to human empathy when a machine is seen as able to supply it. Across four studies (249, 494, 743 and 737 participants) they find a consistent pattern they call empathy offloading: the more capable people believe AI is at providing emotional support, the less willing they are to provide that support themselves, and the effect holds when trait empathy and prior AI use are controlled. Their fourth study complicates the story in a way worth teaching carefully. When the recipient's source of support is named, willingness to offload to AI is lower than willingness to offload to friends or therapists, because people rate AI's empathic capacity far lower. Yet on a harder-to-game behavioral measure the same participants who said they would not offload chose fewer empathy words and more indifference words, and they judged the student who went to AI as stranger and more desperate. The self-report and the behavior disagree, and the paper argues the behavior is where the ethics actually live.

Key Findings

1. Perceived capacity tracks willingness to offload. In Study 1 (N = 249), ratings of ChatGPT's capacity to provide empathy correlated with willingness to offload empathic responsibility to it at b = 0.44, 95% CI [0.35, 0.53], R² = .27, and the relation survived controls for trait empathy and existing use of ChatGPT for emotional support (b = 0.31, 95% CI [0.23, 0.40]). Average belief in ChatGPT's empathic capacity was modest in that sample, 3.24 on a 7-point scale.
2. The link is causal, and runs through belief. Study 2 (N = 494) manipulated perceived capacity with real newspaper articles about AI being used for empathy versus AI generating images. The empathy articles raised willingness to offload (total effect c = 0.83, p < .001), and the effect travelled entirely through perceived capacity: indirect effect ab = 0.73, 95% bootstrap CI [0.52, 0.95], with the direct effect collapsing to c' = 0.10, p = .51. Belief about what the system can do is the mechanism, not exposure to AI as such.
3. Hearing that someone went to AI for support changes behavior. Study 3 (N = 743) told participants an acquaintance in distress had received support from ChatGPT, from another person, or from no one. Willingness to offload was highest after the ChatGPT case (M = 2.79, SD = 1.77) versus no support (M = 2.43, SD = 1.60), p = .04, social-emotional accuracy on the Reading the Mind in the Eyes task dropped in the same comparison (7.13 versus 7.73 correct out of 12, p = .012), and ratings of the person's emotional pain did not differ, so a colder reading of the person's suffering does not explain the effect.
4. Explicit offloading to AI is lower than offloading to people, and the behavior is still worse. Study 4 (N = 737) named the source as friends, a therapist, AI, or no one. Participants rated AI's empathic capacity far below friends and therapists (3.10 versus 5.55 and 5.34, eta-squared = .58) and, consistently, were less willing to offload to AI (M = 2.39) than to friends (3.49) or therapists (3.91). Only in the AI condition did the language-selection task go the other way: those participants chose the fewest empathy words, the most indifference words, and the lowest felt moral duty to help, and they rated the student as more weird, antisocial and desperate.
5. The authors draw four consequences. Empathic AI may devalue empathy by making emotional support look effortless, since people price moral value partly by effort; it may deskill, in the sense of moral deskilling rather than technical deskilling; it may fray the social fabric that ordinary mutual support maintains; and it may let the capacity for empathy atrophy through disuse.
6. Two distinct mechanisms sit inside one label. Studies 1 to 3 support effort-saving offloading, where a believed-capable system licenses less work. Study 4 points at something harsher: withdrawal of empathy as social judgment on the person who sought support from a machine. The authors keep both, and note that an AI-seeking student may lose support twice over.
7. The effect is small in absolute terms and the authors say so. Offloading means across the first three studies ran 2.10 to 3.27 on a 7-point scale, and perceived AI empathic capacity was low throughout, which the paper reports rather than buries.

What this means for practice

• Do not install an emotional-support chatbot and assume the human support stays. This is the one finding an institution can act on directly: everything that raises students' estimate of the system's empathic ability also raises their willingness to hand that work over. If an advising or tutoring service adds an empathic assistant, it should measure whether staff and peer support held steady, not only whether students used the assistant.
• Watch who stops showing up. The wage of offloading is paid in ordinary moments of noticing distress, and a conversation moved to a chatbot leaves no trace in the building. A counseling service that digitizes triage should track how often humans still initiate support.
• Treat AI use for emotional support as stigmatizable, because it currently is. Participants in Study 4 rated a student who went to AI as stranger and more desperate than one who went to a friend or therapist, and selected more indifference words about them. Faculty who normalize the chatbot as the private, shameful option make this worse; framing it as one route among several, alongside people, does not.
• Keep empathy visible as skilled work. The paper's devaluation argument turns on effort: support that appears automatic reads as less morally valuable. Naming what a human listener does, in class and in feedback, is a small counterweight that costs nothing.
• Use it to bound affective tutoring claims. An affective tutor can be genuinely useful to the student in front of it and still thin the surrounding supply of human care. Both statements can be true, and the second does not show up in a satisfaction survey.

Limitations

• Offloading was measured mostly by self-report, and the one behavioral task, a language-selection exercise scored with LIWC dictionaries, is an indirect proxy rather than an observed act of empathy.
• Perceived capacity for AI empathy was low across studies, so the effects sit near a floor. The authors check skewness and kurtosis and rerun the word task with Kruskal-Wallis tests, and the pattern holds, but a population that believed AI more capable might offload more than these estimates suggest.
• Studies 1 to 3 asked participants to imagine a person in distress; only Study 4 named a real student, and that vignette came from one institution.
• All four samples were United States online panel participants recruited through CloudResearch Connect, so cultural variation in the obligation to provide empathy is untested.
• The paper measures willingness and immediate behavior, not the longitudinal claims its discussion makes: devaluation, deskilling, fraying and atrophy are extrapolations from a lab measure, not outcomes it observed.
• Study 4's AI condition included an extra appeal to authority, a cited report on AI use for companionship, which the other conditions did not have, because the authors wanted to overcome low baseline beliefs about AI empathy.

Citation

Shastry, M., & Gray, K. (2026). People offload empathy when they think AI can give it instead. Computers in Human Behavior.