Quantifying Harm

The evolution of AI decision-making has shifted from rule-based systems to complex machine learning models, raising ethical concerns about harm and fairness. Early efforts focused on safety and bias mitigation, exemplified by works like Kleinberg et al. on fairness metrics and Ribeiro’s interpretability methods. As AI impacts critical sectors, the need for systematic harm quantification grew, leading to causal inference frameworks (e.g., Pearl’s structural causal models). Recent studies incorporated decision biases, such as Kahneman and Tversky’s probability weighting, to better model human preferences. However, existing methods often lack a unified approach to quantify harm at both individual and societal levels, especially under uncertainty, limiting their practical utility in policy and ethics.

The main challenge is developing a rigorous, scalable framework to quantify harm that accounts for causal relationships, probabilistic uncertainty, and human biases. Existing models either oversimplify harm as expected utility or ignore biases, leading to inaccurate assessments. Moreover, aggregating harm across populations raises fairness issues, especially when considering vulnerable groups. The difficulty lies in balancing technical rigor with interpretability, ensuring models reflect real-world decision-making processes while remaining computationally feasible. Addressing these gaps is vital for deploying AI systems that are both effective and ethically aligned.

The paper’s key innovations include: 1) formalizing a quantitative harm measure based on structural causal models, ensuring causal interpretability; 2) integrating probability weighting functions (e.g., Prelec’s) to model biases like overweighting small probabilities; 3) proposing social aggregation methods that incorporate fairness constraints, avoiding naive summation biases. These innovations enable nuanced harm assessments that reflect human preferences and social considerations, surpassing prior models limited to qualitative or expected utility frameworks. The approach provides theoretical guarantees and practical algorithms for scalable computation, marking a significant step forward in AI ethics.

• �� Construct structural causal models (SCMs) with endogenous and exogenous variables representing the environment.
• �� Define harm as the utility difference between actual outcome and default outcome, using causal relationships to determine causation.
• �� Incorporate probability weighting functions to adjust for biases in perceived probabilities.
• �� Aggregate individual harms via weighted sums, applying fairness constraints to prevent disproportionate impacts.
• �� Develop algorithms for efficient harm computation, ensuring scalability and robustness across scenarios.

The framework was tested on datasets such as COVID-19 vaccine efficacy and traffic accident probabilities. It compared traditional expected utility models with bias-adjusted models, measuring alignment with human judgments and decision outcomes. Hyperparameters for probability weighting functions were tuned based on empirical studies. Sensitivity analyses examined robustness across different causal structures and population groups. Results showed that the bias-aware models better predicted human preferences and reduced decision errors, especially in low-probability, high-impact events, demonstrating practical relevance.

Quantitative harm metrics correlated strongly with expert assessments, reducing error margins by up to 80%. Incorporating probability weighting improved alignment with human biases, notably overweighting small probabilities in risk scenarios like traffic safety. Fairness-aware aggregation prevented disproportionate impacts on vulnerable groups, maintaining equitable harm distribution. The models scaled effectively to multi-agent scenarios, such as organ transplantation and public health policies, outperforming baseline utilitarian approaches in both accuracy and fairness metrics.

This framework can be directly applied to healthcare decision-making, policy evaluation, and AI system auditing. It enables stakeholders to quantify potential harms precisely, facilitating transparent and ethically informed choices. Future integration with real-time data streams and adaptive bias modeling will support dynamic decision environments, such as personalized medicine and autonomous systems, ensuring ongoing ethical compliance and societal trust.

The approach depends heavily on the accuracy of causal models; errors in causal assumptions can lead to misleading harm estimates. Computational complexity increases with model size, requiring further optimization. Bias models are empirically derived and may not generalize across cultures or contexts. Additionally, the static nature of the current framework limits its application in dynamic, evolving environments. Future work should focus on scalable algorithms, causal discovery, and adaptive bias learning to address these challenges.