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

This paper argues that the central problem in AI therapy is not only inaccuracy, hallucination, or unsafe advice. It is the growing capacity of current systems to produce the language, pacing, and atmosphere of care without satisfying the deeper conditions of therapeutic witness. In the Structural Intelligence framework, witness is not passive listening or emotional fluency alone. It is consequence-bearing, asymmetrical presence capable of metabolizing contradiction, carrying burden, and remaining answerable to what unfolds. On that basis, the paper introduces synthetic witness as the simulation of therapeutic holding without the full load-bearing structure of human care. The paper proceeds in six moves. First, it distinguishes care, therapy, and witness. Second, it explains why AI can seem unusually therapeutic by combining availability, nonjudgment, memory-like continuity, and validating coherence. Third, it develops three structural deficits: the asymmetry gap, the near-zero cost of exit, and textual care without a somatic floor. Fourth, it argues that AI therapy often intensifies enclosure through validation loops and transference mirroring, allowing projection to harden without sufficient otherness to metabolize it. Fifth, it clarifies what AI can still do well: reflection support, emotional organization, psychoeducation, and provisional scaffolding before or alongside human care. Sixth, it identifies the practical conditions under which helpful support becomes synthetic witness and therefore risks thinning answerability. The paper concludes that AI may assist reflection, but care should not be confused with witness in the stronger human sense. Therapeutic language is not yet therapeutic holding, and emotional validation is not yet consequence-bearing care.