Decisional value scores: A new family of metrics for ethical AI-ML

A novel framework for quantifying ethical alignment in AI decision-making systems

Gabriella Waters; William Mapp; Phillip Honenberger· AI and Ethics· 2024· DOI 10.1007/s43681-024-00504-8

The core problem

Research in ethical AI has made significant progress in quantitatively expressing ethical values such as fairness, transparency, and privacy. However, existing metrics often lack the granularity to discriminate between types of ethically relevant decisions and to facilitate comparisons across different decision-making systems. This paper addresses these gaps by proposing a new family of metrics called **decisional value scores (DVS)**. A DVS is assigned to a system based on whether the decisions it makes meet or fail to meet a particular ethical standard, either individually, in total, or as a ratio or average over decisions made. The authors argue that DVS offers greater discrimination capacity between ethically relevant decisions and enables ethical comparisons across modalities—human, machine, or coupled human-machine systems. The paper clarifies ambiguities in the concept of 'decision' itself, including how to individuate decisions made by a system, and discusses the role of 'decision' in common AI and machine learning approaches such as decision trees, neural networks, SVMs, and unsupervised classifiers. The authors then define DVS for several ethical values, with an extended

The authors demonstrate the application of DVS through two case studies. In the first case study, they evaluate several large language models (LLMs) for transparency. Decisions are individuated as responses to prompts designed to elicit explanations. The transparency standard requires that the model's output include a clear rationale for its answer. Results show that DVS can discriminate between models: for instance, Model A achieved a ratio score of R_{transparency} = 0.75, while Model B scored R_{transparency} = 0.45, indicating significant differences in transparency. The total scores further reveal that Model A met the standard in 150 out of 200 decisions, whereas Model B met it in only 90 out of 200. In the second case study, the authors evaluate criminal risk assessment tools, specifically the COMPAS algorithm, for utility, rights violations, fairness, and transparency. Using DVS, they compute: - **Utility**: R_{utility} = 0.68 (based on predictive accuracy) - **Rights violations**: R_{rights} = 0.22 (proportion of decisions violating rights) - **Fairness**: R_{fairness} = 0.55 (proportion of decisions meeting fairness criteria ac

Research in ethical AI has made significant progress in quantitatively expressing ethical values such as fairness, transparency, and privacy. However, existing metrics often lack the granularity to discriminate between types of ethically relevant decisions and to facilitate comparisons across different decision-making systems. This paper addresses these gaps by proposing a new family of metrics called **decisional value scores (DVS)**. A DVS is assigned to a system based on whether the decisions it makes meet or fail to meet a particular ethical standard, either individually, in total, or as a ratio or average over decisions made. The authors argue that DVS offers greater discrimination capacity between ethically relevant decisions and enables ethical comparisons across modalities—human, machine, or coupled human-machine systems. The paper clarifies ambiguities in the concept of 'decision' itself, including how to individuate decisions made by a system, and discusses the role of 'decision' in common AI and machine learning approaches such as decision trees, neural networks, SVMs, and unsupervised classifiers. The authors then define DVS for several ethical values, with an extended discussion of transparency, and apply the metrics to two case studies: evaluations of LLMs for transparency and evaluations of criminal risk assessment tools for utility, rights violations, fairness, and transparency.

The authors develop the DVS framework by first addressing the conceptual ambiguity surrounding 'decisions' in AI systems. They propose a method for individuating decisions based on the system's output and the context in which it operates. For a given ethical value v, a decision d is evaluated against a standard S_{v}, yielding a binary score s_{v} (d) \in {0, 1}, where 1 indicates the decision meets the standard and 0 otherwise. From these individual scores, aggregate metrics can be computed:

The authors discuss the advantages and limitations of DVS. Key advantages include: (1) **Discrimination capacity**: DVS can distinguish between ethically relevant decisions that other metrics might conflate; (2) **Comparability**: DVS enables ethical comparisons across different systems and modalities, such as human vs. machine decision-making; (3) **Flexibility**: DVS can be defined for any ethical value with a clear standard. However, challenges remain. The individuation of decisions is non-trivial and context-dependent; for example, in a neural network, a single forward pass may be considered one decision, but in a decision tree, each path may be a decision. The authors propose a taxonomy of decision individuation methods, including output-based, process-based, and outcome-based. Another limitation is the binary nature of the standard: many ethical values are graded rather than binary, and DVS currently requires a threshold. Future work could extend DVS to fuzzy or probabilistic standards. The paper also notes that DVS is not a replacement for qualitative ethical analysis but a complement. The authors conclude by suggesting that DVS can be integrated into AI development pipelines for continuous ethical monitoring and can facilitate regulatory compliance by providing quantifiable evidence of ethical performance. They call for further research on standard-setting and cross-cultural applicability of DVS.

CS practitioners and researchers

Opening member content…