Large Language Model-Based Chatbots for Depression and Anxiety: Systematic Review and Meta-Analysis (Preprint)

BACKGROUND The global mental health crisis is becoming increasingly severe, with over 280 million patients suffering from depression and over 301 million patients suffering from anxiety disorders. Due to the shortage of mental health professionals, high treatment costs, and insufficient accessibility of services, there is an urgent need for scalable and low-cost intervention methods. Chatbots based on Large Language Models (LLMs) have become a new tool for providing psychological support with advanced natural language processing and deep learning techniques. However, their therapeutic effects are not yet fully studied, especially in terms of differential effects on depression and anxiety. There are obvious limitations to existing research: the psychological assessment tools used in different studies are mixed (e.g., PHQ-9, GAD-7, STAI), and there are significant differences in intervention design (dialogue system architecture, intervention duration), which makes it difficult to directly compare the efficacy evaluation results, especially for the differentiated effects on depression and anxiety. OBJECTIVE This study systematically evaluates the effectiveness of LLM chatbots in alleviating symptoms of depression and anxiety, and analyzes the moderating effects of intervention duration, control group type (e.g., blank control vs. traditional therapy), and demographic characteristics (e.g., age). METHODS By systematically searching PubMed, Cochrane, Scopus, and CNKI, 10 studies (17 effect measures) published between 2020 and 2025 were included, including randomized controlled trials (RCTs) and pre post test designs. The research subjects are adults aged 16 and above with symptoms of depression or anxiety. Using Hedges' g as the measure of effect, a random effects model was employed due to moderate heterogeneity (depression I² = 35.6%, anxiety I² = 54.96%). Subgroup analysis explored the effects of control group type, intervention duration, and age. RESULTS Summary analysis shows that chatbots have a marginal significant improvement effect on depression symptoms (g = 0.157, 95% CI [-0.013, 0.328], p = 0.071), and a significant improvement effect on anxiety symptoms (g = 0.277, 95% CI [0.069, 0.486], p = 0.009). Subgroup analysis found that for depressive symptoms, the effect size of the blank control group (g = 0.309) was higher than that of the traditional therapy group (g = 0.193); For anxiety symptoms, long-term intervention (> 4 weeks) has a significant effect (g = 0.572), far superior to short-term intervention (g = 0.063); Elderly people (> 50 years) benefit more (g = 0.372). CONCLUSIONS LLM chatbots have clinical significance for anxiety, especially in the elderly and long-term interventions, but their improvement on depression is limited. Future research should optimize depression intervention designs, integrate multimodal therapies, and focus on long-term effectiveness.