Abstract
INTRODUCTION: Conversational agents (CAs; computer programs that use artificial intelligence to simulate a conversation with users through natural language) have evolved considerably in recent years to support healthcare by providing autonomous, interactive, and accessible services, making them potentially useful for supporting smoking cessation. We performed a systematic review and meta-analysis to provide an overarching evaluation of their effectiveness and acceptability to inform future development and adoption. AIMS AND METHODS: PsycInfo, Web of Science, ACM Digital Library, IEEE Xplore, Medline, EMBASE, Communication and Mass Media Complete, and CINAHL Complete were searched for studies examining the use of CAs for smoking cessation. Data from eligible studies were extracted and used for random-effects meta-analyses. RESULTS: The search yielded 1245 publications with 13 studies eligible for systematic review (total N = 8236) and six studies for random-effects meta-analyses. All studies reported positive effects on cessation-related outcomes. A meta-analysis with randomized controlled trials reporting on abstinence yielded a sample-weighted odds ratio of 1.66 (95% CI = 1.33% to 2.07%, p < .001), favoring CAs over comparison groups. A narrative synthesis of all included studies showed overall high acceptability, while some barriers were identified from user feedback. Overall, included studies were diverse in design with mixed quality, and evidence of publication bias was identified. A lack of theoretical foundations was noted, as well as a clear need for relational communication in future designs. CONCLUSIONS: The effectiveness and acceptability of CAs for smoking cessation are promising. However, standardization of reporting and designing of the agents is warranted for a more comprehensive evaluation. IMPLICATIONS: This is the first systematic review to provide insight into the use of CAs to support smoking cessation. Our findings demonstrated initial promise in the effectiveness and user acceptability of these agents. We also identified a lack of theoretical and methodological limitations to improve future study design and intervention delivery.