Abstract
<jats:p>Background: The rapid integration of artificial intelligence (AI) into software development has intensified the need for transparent version control practices. However, existing literature lacks a cohesive framework to identify, categorize, and address recurring failures in attributing AI-generated contributions within version control systems (VCS). This fragmented landscape obscures accountability, undermines reproducibility, and complicates compliance with emerging AI governance regulations.</jats:p> <jats:p>Objective: This study conducts a systematic review and meta-analysis to map the prevailing transparency gaps in VCS literature pertaining to AI-assisted or AI-generated code. We aim to synthesize disparate evidence of attribution failures, quantify their prevalence across different development contexts, and propose a structured taxonomy to standardize future reporting and remediation.</jats:p> <jats:p>Methods: Following PRISMA guidelines, we systematically searched Scopus, IEEE Xplore, ACM Digital Library, and arXiv for peer-reviewed and preprint literature published between 2019 and 2026. We included empirical studies, case reports, and tool evaluations that explicitly addressed transparency, provenance, or attribution in VCS workflows involving AI. A random-effects meta-analysis was performed on 47 eligible studies to estimate pooled proportions of reported attribution failure types. Thematic synthesis was subsequently applied to derive a hierarchical taxonomy.</jats:p> <jats:p>Results: Our meta-analysis reveals that attribution failures are reported in 68% (95% CI: 61–74%) of studied AI-integrated projects, with the highest prevalence in commit-message authorship misattribution (42%) and ambiguous provenance of AI-suggested patches (37%). We identify four primary gap clusters: Identity Obfuscation (who contributed), Contribution Boundaries (what was AI-generated), Temporal Discrepancies (when contributions occurred), and Rationale Omission (why changes were made). These clusters inform our proposed Taxonomy of Attribution Failures in AI-VCS Workflows (TAF-AI), comprising 12 distinct failure modes across three strata—individual, tooling, and governance.</jats:p> <jats:p>Conclusions: Current version control transparency mechanisms are ill-equipped for the hybrid human–AI authorship landscape. The TAF-AI taxonomy provides a diagnostic lens for practitioners and a common vocabulary for researchers. We recommend mandatory AI-annotation fields in VCS metadata, real-time provenance tracking, and updated contribution guidelines. Our findings underscore that transparency is not a binary attribute but a multidimensional construct requiring continuous calibration as AI agents evolve from assistants to co-authors.</jats:p>