Abstract
Over the past decade, an ecosystem of measures has emerged to evaluate the social and ethical implications of AI systems, largely shaped by high-level ethics principles. These measures are developed and used in fragmented ways, without adequate attention to how they are situated in AI systems. In this paper, we examine how existing measures used in the computing literature map to AI system components, attributes, hazards, and harms. Our analysis draws on a scoping review resulting in nearly 800 measures corresponding to 11 AI ethics principles. We find that most measures focus on four principles - fairness, transparency, privacy, and trust - and primarily assess model or output system components. Few measures account for interactions across system elements, and only a narrow set of hazards is typically considered for each harm type. Many measures are disconnected from where harm is experienced and lack guidance for setting meaningful thresholds. These patterns reveal how current evaluation practices remain fragmented, measuring in pieces rather than capturing how harms emerge across systems. Framing measures with respect to system attributes, hazards, and harms can strengthen regulatory oversight, support actionable practices in industry, and ground future research in systems-level understanding.
Introduction
The field of Responsible AI (RAI) has seen a proliferation of metrics designed to evaluate the social and ethical implications of machine learning systems. However, these measures are often developed and applied in a fragmented manner. Researchers from McGill University and Google Research argue that we are currently "measuring in pieces"—focusing on isolated technical components while losing sight of how harms actually emerge in real-world sociotechnical contexts. This paper addresses the "measurement problem" by proposing a systems-level approach to AI ethics evaluation.
A Systems-Safety Framework
To bridge the gap between technical metrics and human impact, the authors apply a System Safety lens. In safety engineering, safety is viewed as an emergent property of a whole system rather than a characteristic of individual components. The paper introduces a conceptual chain that maps evaluations from specific system components to ultimate harms.
This framework connects an ML System Component (e.g., Output) to an Attribute Being Measured (e.g., Distribution of system output), which is quantified by a Measure (e.g., Demographic Parity). This measure signals a potential Hazard (e.g., Disparate distribution of output), which may eventually lead to a specific Harm (e.g., Allocative harm).
Methodology
The researchers conducted a comprehensive scoping review of computing literature from 2017 to 2023. They identified 257 papers proposing quantitative or qualitative measures for 11 AI ethics principles. From these, they extracted 791 individual measures and coded them based on the targeted system component (Data, Model, Output, User Interaction, or Full System) and the type of harm signaled (Allocative, Representational, Quality of Service, Interpersonal, and Social System).
Key Findings
The analysis revealed several critical gaps in the current landscape of AI ethics measurement:
1. The "Big Four" Hegemony
Evaluation efforts are highly concentrated. Approximately 90% of all measures focus on just four principles: Fairness, Transparency, Privacy, and Trust. Other vital principles, such as Dignity, Solidarity, and Sustainability, remain largely ignored in technical literature.
2. Model-Centric Bias
There is a significant imbalance in which system components are measured. Over 57% of measures target the model itself, while only 25% examine the interaction between the user and the output. This suggests a technical bias where mathematical properties are prioritized over actual human impact.
3. The Threshold Vacuum
A major hurdle for actionable RAI is the lack of clear "red lines." Only 26.9% of the reviewed measures were accompanied by an explicit criterion or threshold (e.g., the "80% rule" for disparate impact). Without these thresholds, it is difficult for developers to determine when a system state becomes hazardous.
4. Temporal and Interaction Blindness
Most metrics are "one-and-done" snapshots taken during development. They fail to account for how harms accumulate over time or how they emerge from complex interactions across different system elements.
Conclusion
Current AI evaluation practices remain fragmented and disconnected from the lived experiences of harm. To move forward, the authors call for a shift toward system-spanning evaluations. Future research and practice should focus on benchmarks that trace how technical inputs influence model decisions and eventually manifest as human outcomes. Strengthening regulatory oversight and industry practices requires measuring what truly matters: the emergent safety of the full sociotechnical system.
Summary updated