Digitization in forensic document examination: A systematic review of artificial intelligence-based handwriting and signature analysis

Forensic document examination depends heavily on expert opinion, often leading to subjective reports. This review systematically evaluated image processing and artificial intelligence (AI)-based for offline signature and author verification. The evaluation encompasses model architectures, explainability components, and compliance with forensic reporting standards. Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, a systematic review was conducted across PubMed, Web of Science, IEEE Xplore, and Google Scholar for records published until February 1, 2026. Inclusion criteria covered offline signature verification, document preprocessing, and explainability approaches. Online/dynamic signature studies or those with insufficient methodology were excluded. From 200 initial records, strict eligibility assessments yielded 22 studies. Due to methodological diversity, findings were evaluated using a narrative synthesis. Post-2020 studies predominantly rely on AI and transfer learning. Common limitations include non-writer-independent designs, lack of subject-disjoint training/test sets, data leakage risks, and limited external validation. Performance metrics varied significantly, and reporting based on interpretability and likelihood ratio (LR) frameworks remains notably scarce. The synthesis also showed that similarity scores require calibration before they can be interpreted as evidential strength, and that explainability outputs are valuable only when integrated with expert-based forensic reasoning. While AI-based methods show promise for handwriting analysis, significant gaps remain in external validation and reporting standardization. Establishing quantifiable, standardized approaches is urgently required to support expert opinions and ensure the robust forensic validity of evidence in legal proceedings. Future forensic applications should therefore prioritize subject-disjoint validation, calibrated score interpretation, uncertainty reporting, and transparent integration of AI outputs into expert opinion.