Large Language Models as Post-Session Clinical Augmentation Tools: A Theoretical Framework for AI-Assisted Psychotherapy

Psychotherapy is constrained by fundamental limits of human cognition: fatigue, attentional capacity, confirmation bias, and the impossibility of simultaneously tracking both immediate therapeutic process and longitudinal linguistic patterns. This paper proposes the Post-Session Linguistic Augmentation (PSLA) model, a theoretical framework for the use of large language models (LLMs) as post-session clinical augmentation tools that analyse therapy transcripts and return structured, probabilistic, hypothesis-level insights to the treating clinician. Critically, this framework positions LLMs not as diagnostic or therapeutic agents but as perceptual amplifiers operating within a human-in-the-loop architecture. Engaging with Haber et al.’s (2024) conceptualisation of the AI presence in therapy as an ‘artificial third,’ I argue that the risks of AI in psychotherapy are substantially architectural rather than inherent, and that a post-session, clinician-only configuration can access the perceptual advantages of LLM-based linguistic analysis while structurally preserving the therapeutic dyad. The paper details the functional architecture of a five-domain Clinical Augmentation Report, proposes a three-stage validation framework, presents an Uncertainty Management Architecture for handling hallucination and false positive risk, and outlines the ethical requirements governing consent, data sovereignty, and bias. Limitations, contraindications, and a proposed research agenda are discussed.