Designing AI to Prevent Harm (Safety)

The growing integration of artificial intelligence into safety-critical socio-technical systems increases the risk of physical, psychological, and societal harm that cannot be addressed solely by technical reliability. This chapter positions AI safety as a central and independent objective within AI design and governance, emphasizing the prevention, limitation, and containment of harm throughout the entire lifecycle of AI systems. The analysis adopts a harm-based framework and systematically reviews ethical, regulatory, and technical approaches, including the EU AI Act, product safety and liability regimes, safety testing protocols, risks associated with human–AI interaction, and fail-safe mechanisms. The results indicate that adequate AI safety cannot be achieved solely through ex ante testing or formal compliance. Instead, it requires integrated safety architectures that incorporate continuous monitoring, organizational accountability, interaction-aware design, and robust intervention and containment mechanisms in the presence of irreducible uncertainty.