Reading between the lines: LLMs match or exceed human empathic accuracy using text alone

Empathy plays a central role in human emotional relationships. Empathic accuracy, the ability to accurately appraise another person's emotional state, varies by informational modality and, in humans, is often intertwined with emotional and motivational processes. This study examines whether state-of-the-art Large Language Models (LLMs)—GPT-4, Claude, and Gemini—demonstrate accuracy in emotional-appraisal, and how their accuracy compares to that of humans when presented with only the semantic content (transcripts of recorded videos) detailing ecological, complex autobiographical emotional narratives. We compared the emotional-appraisal of LLMs to the empathic accuracy of human participants (N = 127, randomly sampled students, both in-lab and online) who either read the same transcripts or watched the original videos, which enabled them to use facial and bodily expressions, as well as paralinguistic cues, in addition to semantics. LLMs were able to infer emotional states from semantic content alone with a precision that is equal to or surpasses human performance. This was true both generally and when analyzing positive and negative emotions separately. Theoretically, these findings suggest that semantic information alone can support accuracy in emotional appraisal, though humans may not fully leverage this potential. Practical implications are discussed regarding the use of LLMs in introspective and emotional contexts, while raising critical concerns about privacy, ethical risks, and the potential reshaping of emotional understanding, intimacy, and human connection in an increasingly AI-mediated world.