[AI Ethics Seminar EP.2] Research for Properly Understanding the Risks Posed by AI - LG AI Research BLOG

Behind the boundless possibilities of AI technology lies a coexisting concern over side effects. To objectively grasp and prepare for the harm AI may cause, concrete knowledge and experience of actual risks are essential.

Efforts to collect AI incidents occurring around the world continue. Representative examples include the AIID (AI Incident Database), a voluntary record, and the OECD's real-time detection tool, the AIM (AI Incidents Monitor). According to OECD data, the number of AI incidents surged 139% year-over-year between 2023 and 2024, showing a sharp upward trend.

Recently, active efforts have been made to organize scattered incident records into a single system. In particular, MIT built the AI Risk Repository, containing 777 cases, and classified risks according to two criteria.

• Causal Taxonomy: Classification based on the entity behind the risk (human/AI), the timing of occurrence (pre-deployment/post-deployment), and intentionality (intentional/unintentional)
• Domain Taxonomy: Classification into 7 domains including misinformation, malicious use, human-computer interaction, and safety

According to MIT's research findings, the cause of AI risk was most commonly the AI system itself, accounting for 51%, while incidents occurring after deployment accounted for 65%. Such taxonomies can serve as key tools for policymakers' oversight, academic research, and corporations' establishment of risk mitigation strategies.