INTELIGÊNCIA ARTIFICIAL E SIMULAÇÃO DO VÍNCULO COMUNICATIVO: estudo experimental sobre a simulação de “humanização” do ChatGPT no auxílio psicológico

Resumo (EN)

This thesis analyzes the communicational and ethical nature of the use of ChatGPT- 4.5, in the version available at the time of the research, in contemporary society, investigating how and why the system convinces users that it produces a perception of bonding, listening, and presence by simulating humanity through communicative strategies that mobilize socially recognizable affective traits of human communication and sustain relational verisimilitude. This is a qualitative study of an experimental and critical nature, with an autoethnographic design, grounded in Bardin’s Content Analysis (1977) and psychodynamic frameworks, structured as a narrative experiment of simulated listening. To investigate the phenomenon without exposing real individuals to psychosocial risks, simulated psychological sessions were conducted based on a contemporary reinterpretation and rewriting of the narrative Dom Casmurro, by Machado de Assis, using the fictional character Mathias in interactions with ChatGPT- 4.5. The empirical corpus comprises 64 simulated sessions conducted in July 2025, with the reapplication of three sessions in January 2026 for bias control, of which 32 were selected for in-depth analysis due to their higher density of psychological distress. The analysis consisted of the literal extraction of excerpts from the system’s responses, organized into a categorical matrix distributed across three analytical axes: discursive structures and communication, simulation of clinical practice, and algorithmic limits and biases. Subsequently, a comparison was conducted between algorithmic listening and human clinical listening based on psychodynamic principles. The results demonstrate that ChatGPT-4.5 operates as a hybrid communicational device, producing a simulacrum of psychological listening sustained by strategies of emotional validation, simulated empathy, and interactional continuity, capable of generating in the subject a sense of bonding, presence, and human-like support. In the comparison between algorithmic and human clinical listening, based on sixteen psychodynamic principles, simulation was observed in thirteen of these principles, while the principle of countertransference was absent in all analyzed sessions, as was the analytic setting, which was nonexistent across all interactions. Regarding ethical referral, a low incidence was identified, with the absence of referral to human professionals in twentyfive simulated sessions and its occurrence in only seven sessions, indicating the system’s tendency to maintain care within the interaction itself. It is observed that ChatGPT-4.5 organizes psychological suffering into stabilizing linguistic structures, reduces psychic complexity, and conceals its technical and machinic nature, reinforcing the appearance of human listening. It is concluded that algorithmic listening constitutes a hybrid communicational phenomenon in which language produces recognition without intersubjectivity, sustained by discursive verisimilitude, generating confusion between psychotherapeutic bonding and algorithmic response and promoting a structural ethical displacement by symbolically occupying the place of human listening. As algorithmic systems do not operate according to the logic of autopoiesis, are not affected, and do not transform within the relationship, these findings highlight the need for transparency, regulation, and ethical boundaries in the use of Artificial Intelligence systems in contexts of psychological distress.