Troubling Taxonomies in GenAI Evaluation

TL;DR

This paper critiques current societal impact evaluation methods for GenAI, emphasizing the need for context-specific, power-aware models and proposing a governance-first approach to better manage social harms.

Contribution

It highlights limitations of existing taxonomies in societal impact assessments and advocates for a governance-focused framework tailored to social contexts.

Findings

Existing taxonomies are limited in revealing social harms.

Societal impacts are application- and context-specific.

A governance-first approach can better manage harms.

Abstract

To evaluate the societal impacts of GenAI requires a model of how social harms emerge from interactions between GenAI, people, and societal structures. Yet a model is rarely explicitly defined in societal impact evaluations, or in the taxonomies of societal impacts that support them. In this provocation, we argue that societal impacts should be conceptualised as application- and context-specific, incommensurable, and shaped by questions of social power. Doing so leads us to conclude that societal impact evaluations using existing taxonomies are inherently limited, in terms of their potential to reveal how GenAI systems may interact with people when introduced into specific social contexts. We therefore propose a governance-first approach to managing societal harms attended by GenAI technologies.