"Death by a thousand taxonomies?": AI Risk Classification In Practice

An empirical study based on interviews with researchers and practitioners reveals that existing Sociotechnical Outcome Taxonomies (SOTs) for AI risk classification are poorly integrated into practical AI governance processes. This weak integration is attributed to specific features of SOT design and their application, despite the foundational importance of risk classification for regulators, technology firms, and policymakers.

Why it matters

Effective classification of AI risks is fundamental for developing robust governance frameworks and ensuring responsible AI deployment across various sectors. The identified shortcomings in the practical application of risk taxonomies highlight a critical gap that could impede effective regulation, policy-making, and organizational risk management in AI development and adoption.

Key insights

• AI is implicated in a broad range of harms, from unsafe user interactions to societal-wide consequences.
• The classification of AI risks is critical for effective AI governance, informing actions by regulators, technology firms, and policymakers.
• Numerous Sociotechnical Outcome Taxonomies (SOTs) have been developed by researchers and practitioners to structure AI risks.
• SOTs are found to be weakly integrated into current AI governance processes.
• This weak integration is explained by two features related to SOT design and use, though the abstract only specifies 'two features' without detailing them.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.06831