Cross-lingual LLM-mediated affective communication is bounded in fidelity and shaped by prosocial emotion sensitivity

Scalable affective assessment requires generative stimuli that can be tailored to diverse populations, but two prerequisites for individualised use remain untested: whether LLM-mediated affective communication guarantees fidelity of transfer across linguistic context, and whether recipient psychopathology modulates the response to LLM-generated affective stimuli. We introduce a pipeline quantifying the fidelity of LLM-mediated affective communication and testing whether psychopathology modulates responses to LLM-generated stimuli. An English-speaking source cohort (n = 547) provided core affect ratings (valence, arousal, focus, dominance) for 18 emotion constructs. Conditioned on these ratings, GPT-4o generated 180 text and 18 image stimuli, rated by a Korean-speaking evaluation cohort (n = 91) along the same dimensions. Cross-lingual alignment was quantified as cosine similarity and benchmarked against chance, a within-cohort semantic-distinctness threshold (ε), and another same-language cohort (n = 76); associations with six psychopathology measures were tested using linear mixed-effects models. Similarities exceeded chance-level across both formats but did not reach ε, and same-language alignment exceeded cross-lingual alignment. After multivariate adjustment, sensitivity to prosocial emotions (PFQ-2) was the sole predictor of lower-magnitude and more variable responses. LLM-mediated affective communication is bounded and selectively shaped by reactivity to prosocial emotions, providing a measurable channel for individualised affective assessment.