<![CDATA[The integration of large language models (LLMs) and generative AI into mental health services is advancing faster than the process of consolidating clinical evidence and establishing an adequate governance framework. This article presents a narrative review that synthesizes empirical, technical, and conceptual literature on the convergence of LLMs with psychotherapy practice across four main domains: clinical effectiveness, safety and bias, human-AI relational dynamics, and architecture and model deployment. The literature indicates that LLMs possess strong communicative capabilities and are often perceived as having significant potential. However, this potential is not yet supported by an adequate clinical evidence base, as the literature remains dominated by observational designs, scenario-based evaluations, and short-term perception measures. Evidence from controlled clinical trials and longitudinal evaluations that systematically measure changes in clinical symptoms also remains very limited. The limitations of this evidence become increasingly relevant when the use of LLMs is confronted with risks such as crisis assessment, epistemic bias, value mismatches, and relational illusions—all of which indicate that linguistic equivalence does not necessarily reflect therapeutic legitimacy. Based on a cross-domain synthesis, this article proposes a conceptual convergence framework and a risk-stratification-based governance model to evaluate the clinical role of generative systems in psychotherapy. Currently available evidence tends to support the application of LLMs in low-risk contexts and hybrid models with professional supervision, whereas autonomous deployment in high-risk cases still requires stronger clinical evidence to obtain scientific justification]]>
