A Critical Reflection on Large Language Models in Clinical Psychology. Malfunctioning, Metacognition, Memory, and Ethics

Large Language Models (LLMs) are not only technical tools but epistemic actors that simulate understanding, reasoning, and memory. This paper analyses four interrelated dimensions of LLM functioning (i.e., malfunctioning, simulated metacognition, memory, and ethics) through a critical psychological lens. It argues that what appears as reflective or reliable output is often a structurally constrained simulation shaped by linguistic optimization rather than epistemic grounding. Drawing on recent empirical findings, we show how LLMs outperform humans in specific problem-solving tasks without accessing semantic insight, and how linguistic fluency can foster illusions of authority or intentionality. The notion of malfunction is reframed as a systemic byproduct of generative alignment, not as a technical error. Simulated metacognition, while functionally useful, lacks self-representation and introspective depth. Memory, meanwhile, enables narrative coherence but risks anthropomorphic misattribution. Finally, ethics is relocated from the model’s «behaviour» to its design and governance structures. This perspective calls for an epistemology of simulation: one that distinguishes surface competence from cognitive commitment, and that equips researchers and clinicians to critically integrate LLMs into practice without conflating performance with understanding. We also outline two practical tools: the Socratic Stress Test (SST), based on assertion-to-source mapping, principled abstention, and diachronic coherence checks, and a structured functional grid for coding outputs and supporting transparent auditing in clinical contexts while preserving measurement validity across sessions.