AI Chatbots As "Step 0" In Mental Health Care: A Systematic Review and Meta-Analysis of Effects on Human-Human Social Connectedness

Background and Purpose: Stepped-care models aim to match people to the least intensive effective support and escalate when needed. In practice, many people now engage with AI chatbots before contacting services, positioning large language models (LLMs) as a de facto “Step 0” in mental health pathways. This raises an urgent social work question: does AI interaction strengthen or displace human connection in ways that shape disclosure, help-seeking, and movement into care? Because social connectedness is upstream of help-seeking, understanding the interpersonal consequences of chatbot use is essential for ethical stepped-care integration. This systematic review and meta-analysis synthesized quantitative evidence on associations between LLM chatbot use and human-human social connectedness and examined gender as a potential moderator.

Methods: Following PRISMA and MOOSE guidelines, and a preregistered protocol (PROSPERO: CRD42022325540), we searched PubMed, Web of Science, Scopus, and PsycINFO for English-language quantitative studies published from January 2022 onward. Eligible studies assessed interaction with LLM-based chatbots and at least one human-human social connectedness outcome, including loneliness, perceived social support, empathy, communication, or related interpersonal constructs. Two reviewers independently screened studies, extracted data, and assessed risk of bias. Three-level generic inverse-variance meta-analyses were conducted in R to account for multiple effect sizes nested within studies. Experimental/intervention studies and cross-sectional studies were synthesized separately. Meta-regression tested the percentage of male participants as a moderator.

Results: Eight studies met inclusion criteria: four experimental/intervention studies (n = 174) and four cross-sectional studies (n = 3,325), conducted across the United States, China, the United Arab Emirates, and Vietnam. Experimental studies showed a large positive pooled effect on social connectedness outcomes (Cohen’s d = 1.29, 95% CI: 1.06–1.51), suggesting that structured, purpose-built AI interventions can improve specific interpersonal outcomes. However, heterogeneity was high (I² = 70.1%), and leave-one-out analyses indicated that pooled estimates were sensitive to individual studies. Cross-sectional studies showed no significant association between everyday chatbot use and human-human social connectedness (β = -0.10, 95% CI: -0.75 to 0.55), with extremely high heterogeneity. Quality appraisal indicated that experimental studies were good quality and cross-sectional studies were good or moderate quality. Gender did not significantly moderate effects. Publication bias was detected in cross-sectional studies, but trim-and-fill analyses did not change the null finding. No study directly measured help-seeking intentions, disclosure to human supports, transitions from chatbot use to services, or step-up behavior within care pathways.

Conclusions and Implications: Current evidence does not support safe integration of general-purpose AI chatbots into stepped-care frameworks as an assumed beneficial early step. Structured AI interventions may improve empathy, communication, or loneliness in tightly bounded settings, but these findings do not generalize to routine, unstructured chatbot use. For social work research, practice, and policy, the central finding is an evidence gap: without direct measurement of disclosure, escalation pathways, and substitution risk, it remains unknown whether Step 0 AI use augments human connection to care or becomes a low-friction substitute for it. Priority next steps are longitudinal and stepped-care studies that measure social connectedness alongside help-seeking, service uptake, and timely escalation to human support.