Mapping Technologies, Support Functions, Evaluation Outcomes, and Safety Governance Related to Conversational AI Agents for Psychological Support of Patients With Cancer: Scoping Review.

BACKGROUND: Psychological distress and unmet psychosocial needs are common across the cancer continuum. Conversational AI agents may offer scalable support, but their technologies, support functions, evaluation approaches, and safety governance remain fragmented. OBJECTIVE: This review aimed to map the technologies, psychological support functions, evaluation outcomes, and safety governance of conversational AI used by patients with cancer and cancer survivors. METHODS: We conducted a scoping review following Joanna Briggs Institute (JBI) methodology and the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews). Thirteen databases and additional registers were searched through May 5, 2026. Eligible reports examined conversational agents in real or simulated patient-facing oncology contexts with explicit psychological or psychosocial relevance. We summarized system forms and support functions by project and other characteristics by report. No formal critical appraisal was undertaken. RESULTS: We included 24 reports representing 25 study units and 23 underlying projects. Large language model (LLM) or generative AI systems accounted for 39.1% (9/23) of projects. Information or education (23/23, 100%) and symptom or self-management support (n=20, 87%) were the most frequently mapped functions, although these counts included delivered, planned, simulated, co-designed, and participant-reported functions. Structured psychological or behavioral components were delivered in 21.7% (5/23) of projects. An explicit theoretical or conceptual framework was reported in 25% (6/24) of reports. Reports or plans to assess acceptability or user experience (17/24, 70.8%) and engagement (13/24, 54.2%) were more common than reports or plans to assess psychosocial clinical outcomes (8/24, 33.3%) or safety and adverse events (4/24, 16.7%). Human or expert oversight was reported in 62.5% (15/24) of reports, crisis escalation or referral in 33.3% (8/24) of reports, and user-facing AI identity disclosure in 12.5% (3/24) of reports. CONCLUSIONS: Conversational AI research in psycho-oncology spans heterogeneous technologies and stages of development. Clinical and safety evaluation, theoretical foundations, population representation, and oversight responsibilities remain incompletely reported. These findings represent a map of reporting patterns and do not establish effectiveness or safety. Future studies should evaluate psychosocial benefit and safety prospectively and define professional oversight and referral responsibilities. TRIAL REGISTRATION: OSF Registries UM65W; https://osf.io/um65w.