Calls for new metrics, technical standards, and governance mechanisms to guide and evaluate the adoption of ethical Artificial Intelligence (AI) in institutions are now commonplace. Yet, most research and policy eforts do not fully account for all the diferent approaches and issues potentially relevant to the institutional adoption of AI. In this position paper, we contend that this omission stems, in part, from what we call the 'relational problem': the persistence of difering value-based terminologies to categorize and assess institutional AI systems, and the prevalence of conceptual isolation in the fields that study them including ML, human factors, and social science. After developing this critique, we propose a basic ontological framework to bridge ideas across fields-consisting of three horizontal, discipline-agnostic domains for organizing foundational concepts into themes: Operational, Epistemic, and Normative.
prevalence of conceptual isolation in the fields that study institutional AI [
While a myriad ways of categorizing issues related to AI development and adoption have
been proposed, many ultimately rely on unidimensional thinking. That is, they rely heavily on
a single viewpoint or concept—understood both as an abstract idea that ofers a point of view
for understanding some aspect of experience (e.g., bias), and, relatedly, a mental image that
can be operationalized (e.g., measurement bias)—to discuss institutional AI. Much work in the
social and policy sciences stresses the ethical and legal challenges at stake in the adoption of
AI; while research in computer science tends to highlight the computational and operational
aspects that need to be considered. Yet, as pointed out by [
To address this relational gap, we propose a basic ontological framework, described briefly
below, to help bridge terms across fields—consisting of three discipline-agnostic domains for
organizing relevant concepts: Operational, Epistemic, and Normative. Our framework aims to
achieve two key aims: (1) disciplinary reach, i.e., bridge diferent subcommunities (ML, human
factors, social science, policy etc.), and (2) provide impetus for an intellectual shift that reframes
how researchers and key stakeholders (decision-makers, policy creators, advocates, etc.) think
about institutional AI systems. Our framework is ontological in the sense that it is composed of
three simple domains or meta-concepts that aim to act loosely as semantic fields [
Operational Domain The first field is the ‘operational domain’, which aims to represent the topics, issues, and methods related to the routine activities and functionality of institutional AI systems. It is meant to capture terms that are mainly but not exclusively defined, operationalized, and studied in a technical, applied context. More specifically, it is meant to enable researchers to categorize into a single category all relevant concepts that can be employed both as an abstract idea (e.g., ‘accuracy’) and easily operationalized to quantitatively measure a specific attribute of a particular institutional AI system (i.e., ‘percentage of correct predictions’).
Epistemic Domain The epistemic domain aims to capture knowledge-related topics and issues connected to a particular AI system or institutional AI in general. That is, the epistemic domain is meant to help researchers group together concepts that seek to describe properties which pertain to the interface between AI applications and human actors. Both in terms of the knowledge, beliefs, and intentions of those using AI applications (e.g., a desire for transparency), and the internal properties of the system itself (e.g., its interpretability).
Normative Domain The meaning and uses of concepts in the normative domain, the final
domain we propose, collectively relate to the entitlements, values, and principles of political
morality that stakeholders and afected parties hold towards a particular algorithmic system
or institutional AI in general. The term ‘political morality’ is used here to refer to normative
principles and ideals regulating and structuring the political domain [
Taken together, the framework’s utility derives from the fact that it is discipline-agnostic. More specifically, it aims to be instructive for the individual researcher studying institutional AI in both helping with organizing concepts used to study AI systems and, perhaps more importantly, by drawing attention to whether all potential topics—by virtue of being relevant to one or more of the three proposed domains—have been accounted for. Overall, our contribution
All institutional AI Set of applications Single application
Routine machine activities and technical functionality
Knowledge-related issues at human-machine interface
Moral entitlements, principles and values of affected groups § Accuracy § Efficiency § Reliability §§ [R…o]bustness § Metrics: ‣ Classification accuracy § Standards: ‣ Performance benchmarks ‣ Reporting cards […] § Explainability § Interpretability § Reproducibility §§ [T…ra]nsparency § Mechanisms: ‣ Human-in-the-loop § Metrics § Standards: ‣ Documentation checklists § […] § Equality § Fairness § Priority § Welfare § […] § Mechanisms: ‣ Algorithmic impact
Assessments § Metrics: ‣ Classification parity § […]
Contingent on application context aims to benefit the algorithmic fairness community by facilitating a constructive dialog around the challenges we face as a diverse, interdisciplinary field.