Beyond Behaviorist Representational Harms: A Plan for Measurement and Mitigation

Abstract: Algorithmic harms are commonly categorized as either allocative or representational. This study specifically addresses the latter, focusing on an examination of current definitions of representational harms to discern what is included and what is not. This analysis motivates our expansion beyond behavioral definitions to encompass harms to cognitive and affective states. The paper outlines high-level requirements for measurement: identifying the necessary expertise to implement this approach and illustrating it through a case study. Our work highlights the unique vulnerabilities of LLMs to perpetrating representational harms, particularly when these harms go unmeasured and unmitigated. The work concludes by presenting proposed mitigations and delineating when to employ them. The overarching aim of this research is to establish a framework for broadening the definition of representational harms and to translate insights from fairness research into practical measurement and mitigation praxis.

• The paper redefines representational harms by integrating both affective and cognitive dimensions into its measurement framework.
• It operationalizes harm detection by translating theoretical definitions into practical, interdisciplinary metrics that involve community collaboration.
• The paper proposes innovative mitigation strategies for LLM-induced harms, including seamful design and counter-narratives to challenge societal biases.

Introduction

Algorithmic systems today stretch across critical sectors, affecting decision-making in areas as sensitive as employment, finance, and scientific research. Their benefits are, however, juxtaposed with potential harms, particularly in the field of fairness. This paper presents a deep dive into the often-overlooked category of algorithmic harms known as representational harms, which differ from the more frequently discussed allocative harms. While allocative harms relate to material disparities, representational harms pertain to more intangible damages that impact social standings, cognitive states, and cultural cohesion. The authors argue for an amplified focus on these harms, especially due to the exponential growth and capabilities of LLMs.

Representational Harms: Definition and Extent

The prevailing literature primarily defines representational harms through observable behaviors, ignoring the internal cognitive and affective impact such harms may have. This focus on exterior manifestation limits the understanding of the potential damages of these harms. The authors of the paper conducted a review revealing most harm definitions are behaviorist and reliant on correlations that may not capture the true causative nature of representational harms. By expanding the definition to include affective and cognitive changes, they underscore the broader, more nuanced implications of such harms when inflicted by LLMs, which are at particular risk due to their expansive reach in everyday applications.

Operationalizing Harm Detection

The paper moves from theory to practice, discussing the translation of expanded harm definitions into real-world measurement and mitigation. Implementing practical frameworks demands interdisciplinary expertise and community engagement, given the complexities of capturing representational harms at both individual and collective scales. The authors point out the inherent difficulties of measuring these harms, balancing the need for precision with the practical limitations of deploying surveys and passive measures that can accommodate diverse user demographics and varying cultural contexts.

Unique Challenges with LLMs

LLMs, due to their sophisticated, often seamless design, raise unique concerns regarding the propagation of representational harms. The authors highlight the anthropomorphic interactions and normative outputs of LLMs that may inadvertently entrench societal biases in their normative presentation of information. Coupled with the ubiquity of LLMs in a multitude of platforms and interfaces, these features can lead to a silent but pervasive spread of representational damages that evade user detection and critical scrutiny.

Mitigation Strategies

Finally, this work recommends mitigation strategies, emphasizing that current approaches are inadequate and highlight innovative possibilities like seamful design to instigate mindfulness in users, and counter-narratives to diversify and deepen representations. Emphasizing the importance of careful measurement and discernment in harm propagation, the paper also acknowledges that there may be morally defensible cases where some representational harms, while not ideal, might be a necessary component of transitioning to a more equitable status.

Conclusion

This examination of representational harms serves as a clarion call to researchers, developers, and policymakers to acknowledge and tackle the complex webs of harm that extend beyond the physical and financial into the arenas of cognition and emotion. The proposed framework for harm measurement and mitigation sets a foundation for future AI applications to contend with the multifaceted impacts of their technologies on society.