From Wittgenstein’s Language Games to LLMs: Representation, Meaning, and the Future of Psychotherapy

Abstract

AI and large language models (LLMs) are rapidly taking center stage in clinical psychology and psychotherapy, raising fundamental questions about how machines represent language, emotion, and lived experience. LLMs exhibit linguistic fluency and contextual adaptability, but these capabilities are based on statistical relationships rather than human cognition, presenting significant limitations in therapeutic contexts. Drawing on Wittgenstein's concepts of language games and forms of life, this article examines representation problems in AI-assisted therapeutic dialogue and argues that LLMs can imitate but not genuinely internalize the relational and experiential dimensions fundamental to psychotherapy. The analysis explores how emotional simulation varies across LLMs, why simulated empathy diverges from therapeutic empathy, and how clinical safety concerns arise from limitations in contextual reasoning and emotional attunement. Based on a focused narrative review of 36 studies from 2015–2025, findings indicate that LLMs offer potential for assessment, diagnostic support, training, and psychoeducation, but remain limited in representing affect, cultural nuances, and embodied co-regulation essential to therapeutic relationships. The authors propose alternative systems based on hybrid neuro-symbolic architectures, multimodal affect systems, and interdisciplinary collaboration.