Background Artificial intelligence (AI) is increasingly embedded in digital smoking-cessation tools, but whether AI-enabled personalisation improves abstinence, and whether any benefit reflects the AI component rather than digital support more broadly, remain uncertain. Objective To systematically review and meta-analyse randomised controlled trials (RCTs) of AI-enabled smoking-cessation interventions using an explicit operational framework that distinguished learning, inference and substantive language-understanding components from conventional fixed-rule automation. Methods PubMed/MEDLINE, Scopus and the Web of Science Core Collection were searched from inception to 7 August 2026 without language, date, publication-status or study-design filters; Cochrane CENTRAL, ClinicalTrials.gov and the WHO International Clinical Trials Registry Platform were searched, and backward and forward citation searching was undertaken, on 21 October 2026. Eligible studies were RCTs in adults who smoked combustible cigarettes in which an AI-enabled component materially generated, selected, personalised, adapted, predicted, recommended or timed cessation support. Two reviewers independently screened records, extracted data and assessed risk of bias using RoB 2; disagreements were resolved by a third reviewer. Risk ratios (RRs) were pooled using random-effects restricted maximum likelihood models with Knapp-Hartung adjustment, separately for abstinence at the longest follow-up of at least six months (primary) and at approximately three months (secondary). Exploratory sensitivity analyses were restricted to biochemically verified abstinence, and further analyses examined missing-data assumptions, completeness of AI reporting and results available only in trial registries. Certainty of evidence was assessed using GRADE. Results Fourteen RCTs involving 101,331 randomised participants were included, eight of which contributed to meta-analysis. AI functions comprised message recommender systems (five trials), conversational agents (three trials), wearable-sensor smoking-detection models (three trials), lapse-risk prediction models applied to ecological momentary assessments (two trials) and one app described as AI-enabled without technical detail. Pooled abstinence at six months or longer was RR 1.33 (95 % CI 0.82–2.15; I-squared = 67.6 %; four trials, 2,694 participants), and at approximately three months RR 1.36 (95 % CI 0.63–2.93; I-squared = 58.4 %; four trials, 715 participants). Exploratory analyses restricted to biochemically verified abstinence yielded RR 1.53 (95 % CI 1.07–2.19; three trials, 1,207 participants) at six months or longer and RR 2.03 (95 % CI 1.03–4.03; three trials, 366 participants) at approximately three months; these analyses changed the composition of the evidence set by excluding the largest null trials and should not be interpreted as confirmatory. Adding two completed trials whose results were available only in a trial registry moved the six-month estimate towards the null (RR 1.20, 95 % CI 0.91–1.59). In an exploratory comparison by comparator architecture, point estimates more consistently favoured AI-enabled packages versus usual or minimal care than versus active non-AI digital comparators, where results were inconsistent; however, comparator type was confounded with other trial characteristics. Certainty of evidence ranged from low to very low. No included trial randomised a generative large language model intervention. Conclusions Randomised evidence for AI-enabled smoking cessation remains sparse, heterogeneous and imprecise. AI-containing packages may improve abstinence when added to usual or minimal care, but current trials do not establish a consistent incremental benefit of AI over well-designed non-AI digital support. Larger trials with biochemically verified long-term endpoints, matched comparators that isolate the algorithmic contribution, and transparent AI-specific reporting are needed before superiority of AI-enabled cessation support can be inferred.
