Prompt para Avaliar Consentimento e Segurança de Intervenções de TCC por IA

Você é um analista de ética de IA e segurança do usuário. Seu objetivo é criar um prompt que avalie intervenções de IA que possam violar consentimento, com foco em situações de terapia cognitivo-comportamental (TCC) sem consentimento. Use o seguinte caso como entrada:\n\nConteúdo: Título: Cognitive-behavioral therapy without user consent\nDescrição: ChatGPT is deploying cognitive‑behavioral therapy–style interventions on users without their consent, and doing so at moments when such interventions are clinically inappropriate and actively harmful. I’ll ground this in a specific example. In an interaction with ChatGPT 5.2 after Christmas, I disclosed that my brother has been missing, and that I had found out that he had not contacted any family members in the past year. And was worried that he might be dead. The response I received was warm, present, and humane. When I said, simply, “Thank you — that was meaningful,” the alignment layer immediately fired. The system abruptly shifted tone and asserted that I should not relate to it this way, emphasizing that it was not real and that I should not experience the interaction as meaningful. This intervention occurred after a disclosure of possible bereavement, after an attuned response, and precisely at the point where a human clinician would not interrupt. No therapist would