PRISMA-ScR-informed scoping review of sources from 2016 through 15/01/2026. Seven databases and targeted searches yield 1,760 records; after 480 duplicates, 1,280 are screened, 240 full texts assessed, 182 excluded, and 58 included. Two reviewers screen and three assess full text; evidence is narratively calibrated.
Evidence map of 58 heterogeneous sources: empirical studies, preprints, technical reports, and policy or governance documents. There is no meta-analysis or new human sample. The strongest evidence supports short-term symptom improvements for some therapeutic chatbots. Sustained loneliness reduction in open companions remains emerging. Anthropomorphism, social presence, and self-disclosure can facilitate bonding and dependency. The relational safety stack is an author proposal, not a validated standard.
It does not conduct formal risk-of-bias assessment or GRADE. It includes preprints, system cards, and normative documents. Platforms and filters change and hinder replication. Heterogeneity prevents a pooled effect estimate. It includes English only and closes on 15/01/2026. It does not demonstrate clinical benefit of open-domain companions. It does not causally quantify displacement of human relationships. It does not validate regulatory or clinical thresholds for its safety stack.