Semantic Classification of Artificial Intelligence Incidents Based on Vector Embeddings and the TAIM Framework

Abstract

The article examines the semantic classification of artificial intelligence incidents based on vector embeddings and the TAIM framework. The aim of the study is to develop and validate a methodology that links textual descriptions of AI incidents with the Govern, Map, Measure, and Manage domains. The relevance of the work is determined by the growing number of incidents, the fragmentation of data sources, and the limitations of manual expert annotation. The scientific novelty lies in the construction of an end-to-end pipeline that transforms unstructured reports into quantitative metrics of semantic similarity with TAIM control areas, while preserving confidentiality through local use of Ollama. It is shown that vector embeddings increase the completeness of identifying conceptually related risks, including cases of terminological divergence. The largest share of incidents falls within the Manage domain, which indicates the importance of infrastructural, organizational, and procedural response measures. The obtained results confirm the potential of semantic mapping for AI audit, monitoring, and risk management. The article will be useful for specialists in AI Governance, information security, compliance, AI auditing, and researchers of digital technology risks.