Abstract
Understanding the landscape of potential harms from algorithmic systems enables practitioners to better anticipate consequences of the systems they build. It also supports the prospect of incorporating controls to help minimize harms that emerge from the interplay of technologies and social and cultural dynamics. Based on a scoping review of computing research (n=172), we present an applied taxonomy of sociotechnical harms to support a more systematic surfacing of potential harms in algorithmic systems. Five major themes related to sociotechnical harms—representational, allocative, quality-of-service, interpersonal harms, and social system/societal harms—and sub-themes are presented along with a description of these categories.
Introduction
As algorithmic systems transition from controlled laboratory environments to real-world applications, they frequently produce "sociotechnical harms"—adverse experiences emerging from the complex interplay between technical code and social power dynamics. Currently, the terminology used to describe these harms is fragmented across disciplines; for instance, an engineer might identify "performance disparities" while a sociologist notes "minoritizing practices." This lack of a synthesized overview hinders practitioners from systematically anticipating and mitigating the consequences of the systems they develop. This paper addresses this gap by providing a comprehensive taxonomy of sociotechnical harms.
Methodology
The researchers conducted a scoping review and reflexive thematic analysis to map existing literature. Following the PRISMA-ScR framework, they curated a corpus of 172 articles and frameworks from major computing databases and professional organizations. Two researchers independently coded the data, clustering specific harms into broader themes. This iterative process resulted in a hierarchical taxonomy that anchors canonical concepts in responsible machine learning while expanding into interpersonal and societal impacts.
The Taxonomy of Sociotechnical Harms
The primary contribution of this work is the Applied Taxonomy for Harm Reduction, which organizes sociotechnical harms into five major pillars.
- Representational Harms: How systems reinforce unjust hierarchies through stereotyping, erasure, or demeaning depictions.
- Allocative Harms: How systems unfairly distribute resources, such as housing, employment, or credit.
- Quality-of-Service (QoS) Harms: Performance disparities where a system functions less effectively for certain groups (e.g., voice recognition failing for specific dialects).
- Interpersonal Harms: How technology shapes relations between individuals, including tech-facilitated violence and loss of agency.
- Social System Harms: Macro-level impacts on democracy, culture, and the environment.
These harms are often interdependent. For example, a representational harm, such as stereotyping a group as "high risk," can lead directly to an allocative harm, such as the denial of a loan.
Granularity and Examples
The taxonomy is designed to be "cascading," moving from high-level themes to granular sub-types with concrete examples. This structure allows policy makers to use high-level categories for regulation while enabling engineers to operationalize specific tests for evaluation.
The review found high consensus in representational and allocative harms but identified social system and interpersonal harms as under-researched areas lacking standardized definitions.
Discussion and Limitations
While the taxonomy provides a systematic framework, it faces certain limitations. The corpus primarily consists of English-language academic literature, which may miss harms experienced in the Global South or those framed through non-Eurocentric worldviews. Additionally, there is a "taxonomy paradox": while a list helps systematic analysis, it may also limit the imagination of practitioners, leading them to "check boxes" rather than critically considering novel, emergent harms.
Conclusion
This five-part taxonomy of sociotechnical harms provides a shared vocabulary for AI practitioners to move from vague concerns about "bias" to specific, actionable risk assessments. By identifying how algorithmic systems impact individuals and society, this framework supports the development of more responsible and equitable technologies.
Summary updated