General-Purpose LLM Chatbots as Informal Step-0 Support: Systematic Review and Meta-analysis of Human–Human Social Connectedness Outcomes (Preprint)
Bibliographic record
Abstract
BACKGROUND General-purpose artificial intelligence (AI) chatbots powered by large language models (LLMs) are increasingly used as always-available sources of companionship, advice, and emotional support. In real-world practice, this pattern can position everyday chatbots as a de facto “step 0” within emerging stepped-care ecosystems. However, the evidence base has emphasized symptoms and usability more than interpersonal processes, and it remains unclear whether AI-human interaction strengthens or erodes human–human social connectedness (eg, loneliness, perceived social support, interpersonal communication, empathy), outcomes that are plausibly proximal to disclosure and downstream help-seeking. OBJECTIVE This systematic review and meta-analysis aimed to synthesize quantitative evidence on the association between LLM chatbot interaction and human–human social connectedness across populations and study designs, and to identify measurement gaps that constrain inference about safe stepped-care integration. METHODS We conducted a PRISMA- and MOOSE-guided systematic review (PROSPERO registered) of studies published from January 1, 2022 onward. Searches were conducted in PubMed, Web of Science, Scopus, and PsycINFO, with an updated search in June 2025. We included quantitative and mixed-method/intervention studies in which LLM chatbot interaction was the primary exposure and social connectedness–related constructs were outcomes. Studies of embodied agents and legacy voice assistants were excluded. Two reviewers screened and extracted data, and risk of bias was assessed using the Newcastle-Ottawa Scale. We conducted three-level generic inverse-variance meta-analyses to pool effects separately for experimental/intervention studies (Cohen d) and cross-sectional studies (standardized β), with meta-regression testing moderation by participant sex (percentage male). RESULTS From 5302 records, 8 studies were eligible for meta-analysis (4 experimental/intervention; 4 cross-sectional). Experimental/intervention studies (174 participants) showed a large positive effect of AI-human interaction on social connectedness outcomes (Cohen d=1.29, 95% CI 1.06-1.51; k=11 effect sizes), with substantial heterogeneity (I²=70.1%). Cross-sectional studies (3325 participants) showed no statistically significant association between LLM chatbot use and social connectedness (β=-0.10, 95% CI -0.75 to 0.55; k=12), with extreme heterogeneity (I²=99.8%). Sex did not significantly moderate effects. Sensitivity analyses indicated that the pooled experimental effect was highly contingent on individual studies. CONCLUSIONS Structured, purpose-built AI-supported interactions can improve social connectedness–related outcomes under controlled conditions, but current evidence does not support assuming that every day, naturalistic chatbot use reliably enhances real-world human connectedness. Crucially for stepped-care framing, none of the included studies assessed help-seeking intentions or behavior, nor transitions from chatbot use to human sources of support, limiting inference about escalation, substitution, or safe “step 0” implementation. Future work should prioritize longitudinal and stepped-care designs with standardized interpersonal outcomes and validated help-seeking/escalation measures to determine whether everyday AI use augments human support or increases substitution risk. CLINICALTRIAL PROSPERO registration number: CRD42022325540
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.030 | 0.119 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.011 | 0.019 |
| Bibliometrics | 0.009 | 0.010 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.005 | 0.003 |
| Open science | 0.002 | 0.003 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.007 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".