New Taxonomy Maps Workplace AI Agent Risks from 2,078 Job Tasks
A recent paper published on arXiv presents a comprehensive multi-layer framework along with a 15-category taxonomy designed to categorize workplace risks associated with AI agents. This investigation examined 2,078 job tasks sourced from the O*NET database, resulting in the creation of 8,356 risk scenarios classified by severity and mode of deployment, whether automation or augmentation. Validation of these scenarios involved 45 workers across 10 different job roles and an independent LLM judge. This taxonomy fills a significant void in current AI risk classifications, which often overlook job-specific hazards. The framework encompasses three essential components: agents, goals, and environment, aiming to assist organizations in predicting socio-technical risks posed by AI agents.
Key facts
- The paper is available on arXiv with ID 2608.08601.
- The study developed a multi-layer framework from a literature review of AI agents.
- The framework models agents, goals, and environment.
- The framework was applied to 2,078 job tasks from the O*NET database.
- The analysis produced 8,356 risk scenarios.
- Risk scenarios were labeled by severity and deployment mode (automation or augmentation).
- Validation involved 45 workers across 10 job roles and an independent LLM judge.
- The resulting taxonomy has 15 categories of workplace AI agent risks.
Entities
Institutions
- arXiv
- O*NET database