Abstract
<p>People increasingly turn to AI chatbots for emotional support, and a growing body of work suggests that these systems can produce responses that are rated as more empathic than those of highly trained humans. Far less attention has been paid to whether these empathic responses are also relationally appropriate, meaning whether they maintain suitable boundaries, avoid fostering dependency, and support rather than substitute for users' own coping and social resources. The present study separates perceived empathy from relational appropriateness and tests whether these two dimensions diverge across large language models. Forty naturalistic, emotionally vulnerable prompts were submitted to the flagship model from four major providers (Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and Llama 4 Maverick), yielding 160 responses that were rated by two trained coders on validated empathy items and a novel relational appropriateness scale developed for this work. Perceived empathy and relational appropriateness were both highly reliable and unidimensional, but they related to one another differently across models (F(3, 112) = 4.39, p = .006). For most models, higher perceived empathy was associated with higher appropriateness, but for one (Llama 4 Maverick) this relationship was flat to slightly negative. Qualitative examples illustrate how a response can be rated as highly empathic while simultaneously discouraging external help-seeking and positioning the AI as a relational substitute. These findings suggest that empathy and relational appropriateness are separable constructs that should be evaluated independently, and that psychologists should be involved in the design of empathic AI safeguards.</p>