The paradox of defining what machines must learn to feel

Abstract

This thesis examined how artificial intelligence can detect and generate empathy in written language to support service design, especially for small and mediumsized enterprises (SMEs). Its purpose was to clarify what “empathetic AI” currently means, where the field stands, and which research gaps limit practical deployment. The work set three objectives: (1) map the spectrum of definitions and operationalisations of empathy used in AI research, (2) review technical and methodological approaches for identifying or producing empathic text, and (3) assess the implications of these approaches for ethically sound, user-centred service design. A scoping-review strategy following Arksey and O’Malley’s fivestage framework was applied. Relevant literature (2017–2024) was retrieved, then filtered through predefined inclusion and exclusion criteria; twenty-seven peer-reviewed studies met the criteria. The review shows that empathy in AI is defined inconsistently. Transformer-based language models and synthetic-data augmentation have improved performance, yet domain adaptation, interpretability and cultural sensitivity remain open problems. A consensus is emerging that future systems need modular architectures combining emotion recognition, perspective-taking and adaptive response generation, supported by transparent evaluation frameworks. For practitioners, the findings offer design principles for embedding low-cost, empathic text tools into SME service workflows while avoiding superficial or manipulative interactions. Researchers may use the identified gaps as a springboard for interdisciplinary collaboration and benchmark creation.