Automatic Classification of Records and Archives as Data: A Survey of Experiments Using Machine Learning

This paper investigates the application of machine learning (ML) to the automatic classification of records and archives, framing it as a critical challenge in Knowledge Organization (KO). As digitization creates massive volumes of uncategorized data, the following research question arises: how can fundamental archival principles—such as provenance, original order, and hierarchical description—be translated into this new computational paradigm? This study first synthesizes, based on a multidisciplinary review of archival science, classification theory, KO, computer science, and information science, a proposal of six fundamental guidelines for the responsible application of artificial intelligence (AI) in records and archives. These guidelines connect traditional archival theory with the modern imperatives of trustworthy and explainable AI. Second, we conduct a comparative analysis of 24 published ML experiments, assessing their adherence to these guidelines. Our analysis reveals a significant and troubling disconnect. While most experiments acknowledge the principle of provenance (75.0%), they demonstrate profound neglect of guidelines related to diverse perspectives (25.0%), explainability (16.7%), and, most critically, algorithmic accountability (0.0%). The results indicate that current practices often succeed in basic content categorization but fail in the more sophisticated archival task of preserving archives’ evidentiary and relational integrity by treating records as decontextualized data. The study calls for urgently developing an Archival AI Lifecycle—a framework that weaves archival principles, classification theory, and knowledge organization into AI development, safeguarding archival practice’s intellectual and ethical integrity in the digital age.