AI Therapy, Care, and Synthetic Witness: Structural Intelligence and the Limits of Machine Care

Conversational AI can provide psychoeducation, structured exercises, emotional organization, and relationship-like interaction, and growing evidence suggests that some mental-health chatbots can improve symptoms or support engagement for some users. These effects should not be dismissed, but neither should they be equated automatically with clinician-delivered psychotherapy. This paper develops synthetic witness as a bounded design-risk concept for cases in which care-like conversational cues are interpreted as stronger clinical, relational, or accountability capacities than the system actually possesses. The concept does not deny the possibility of a digital therapeutic alliance: recent research finds that users can experience goal alignment, task agreement, trust, bond, and continuity with AI-mediated tools. The narrower question is what those experiences authorize. The paper distinguishes perceived alliance from clinical competence, relational effect from reciprocal personhood, and conversational continuity from professional accountability. It rejects embodiment or bodily co-regulation as a universal requirement for therapy, since effective care can occur through telehealth and other mediated forms. Instead, the critical boundaries are role clarity, evidence of efficacy, crisis handling, privacy, informed consent, continuity, human escalation, and responsibility for adverse outcomes. AI support is therefore best evaluated by what it reliably does, under what safeguards, and for which populations, rather than by whether its language feels therapeutic.