Computer vision in transitional justice informatics: A review of technologies, ethics, and forensic applications

Autor/es

Hugo Rojas

Domingo Mery, Jocelyn Dunstan

Resumen

Human rights investigators are frequently overwhelmed by massive, physically degraded volumes of visual evidence. This review evaluates the technological frameworks driving computational forensics in transitional justice. By analyzing multimodal deep learning methodologies, spanning remote sensing, forensic anthropology, text recognition, and biometric identification, we demonstrate how specialized artificial intelligence (AI) pipelines extract legally actionable intelligence from highly fragmented archives. However, our findings highlight acute evidentiary risks, particularly demographic algorithmic bias and the legal inadmissibility of generative AI hallucinations as substantive standalone evidence. Consequently, our primary contribution is the synthesis of a legally defensible operational framework. To satisfy rigid judicial admissibility thresholds, human-in-the-loop methodologies must be implemented not merely as a technical safeguard, but as a mandatory forensic protocol. We ground this framework in a case study of the NuestramemorIA project, utilizing AI to explore the Chilean dictatorship (1973–1990) archives, ultimately illustrating how combining computational scale with independent human validation establishes robust evidentiary links for legal accountability.