AI Safety: Rethinking Risk Beyond Just the Hazard

The article challenges the conventional approach to AI safety, arguing that risk is not solely a property of the AI model's capabilities (hazard) but also depends on exposure and vulnerability. It uses the analogy of a toxic waste dump to illustrate how focusing solely on reducing the hazard (toxicity) can be inefficient if exposure is minimal. The author proposes a framework where risk is a function of hazard, exposure, and vulnerability, suggesting that reducing any of these factors can lower overall risk. The article applies this framework to the AI labor market disruption, noting that while the hazard (AI capabilities) is increasing, policymakers can address risk by managing exposure (AI adoption) and vulnerability (economic reliance on white-collar jobs). The author suggests that policymakers might have limited control over the hazard but can influence exposure and vulnerability through policies and regulations. The article advocates for a more nuanced and practical approach to AI safety, emphasizing the importance of considering the broader context and the interplay of various factors. This perspective encourages a shift from solely focusing on technical alignment to also addressing societal and economic vulnerabilities.