Systematic review on the effectiveness of AI-delivered CBT interventions for anxiety symptoms

Introduction: Cognitive Behavioral Therapy (CBT) is a well-established psychotherapeutic approach for treating anxiety symptoms, leveraging cognitive restructuring and behavioral interventions to modify maladaptive thought patterns. With advancements in artificial intelligence (AI), it has emerged as a promising tool to deliver CBT-based interventions. These AI-driven systems utilize machine learning, algorithms, natural language processing (NLP), or customized messages to personalize interactions and provide real-time therapeutic guidance. This systematic review examines the effectiveness of various AI-delivered CBT interventions for reducing anxiety symptoms. Methods: Following PRISMA guidelines, two systematic searches on Web of Science, PubMed, and PsycINFO were conducted on November 6, 2024 and 13th April, 2025. Eligibility was determined using the PICOTS framework, including only peer-reviewed randomized controlled trials (RCTs) assessing AI-based CBT interventions for anxiety symptoms. Quality assessment measures were implemented to ensure trial's quality. Results: Out of 14 studies, only 1 showed anxiety reduction using AI CBT in an intention-to-treat analysis and 5 showed effects in a completer analyses. Interventions using third-wave CBT (like ACT or mindfulness) were more effective than second-wave CBT. AI CBT worked better for younger groups while all studies on adults 50+ were ineffective. More advanced AI (e.g. generative models, more personalized responses) wasn’t more effective than simpler AI (e.g.rule-based AI, less personalized responses). Adding human support did not improve outcomes either. Limitations: Most studies had at least some risk of bias, which are mainly due to high dropout rates and weak study designs. Many lacked pre-registered plans and only used a per-protocol analysis. Differences in AI design, delivery, and CBT content made comparisons difficult, highlighting the need for standardized reporting of AI interventions. Conclusions: This review encourages researchers to consider which types of CBT interventions work best with AI delivery, who benefits the most from AI-delivered CBT interventions, and whether personalization, improved technology, or human intervention is necessary for better outcomes. More high-quality studies are needed to fully understand the effectiveness of AI-delivered CBT interventions on anxiety outcomes.