Recourse, Repair, Reparation, & Prevention: A Stakeholder Analysis of AI Supply Chains

TL;DR

This paper analyzes AI supply chains to identify stakeholders, harms, and responses, proposing a typology of remedies to promote responsible AI deployment across different market structures.

Contribution

It introduces a stakeholder analysis framework and a typology of responses to harms in AI supply chains, emphasizing design considerations for responsible AI management.

Findings

Stakeholders face diverse harms depending on their role in AISCs.

Market dynamics influence the type and likelihood of remedies implemented.

A typology of recourse, repair, reparation, and prevention guides responsible AI practices.

Abstract

The AI industry is exploding in popularity, with increasing attention to potential harms and unwanted consequences. In the current digital ecosystem, AI deployments are often the product of AI supply chains (AISC): networks of outsourced models, data, and tooling through which multiple entities contribute to AI development and distribution. AI supply chains lack the modularity, redundancies, or conventional supply chain practices that enable identification, isolation, and easy correction of failures, exacerbating the already difficult processes of responding to ML-generated harms. As the stakeholders participating in and impacted by AISCs have scaled and diversified, so too have the risks they face. In this stakeholder analysis of AI supply chains, we consider who participates in AISCs, what harms they face, where sources of harm lie, and how market dynamics and power differentials inform the type and probability of remedies. Because AI supply chains are purposely invented and implemented, they may be designed to account for, rather than ignore, the complexities, consequences, and risks of deploying AI systems. To enable responsible design and management of AISCs, we offer a typology of responses to AISC-induced harms: recourse, repair, reparation or prevention. We apply this typology to stakeholders participating in a health-care AISC across three stylized markets $\unicode{x2013}$ vertical integration, horizontal integration, free market $\unicode{x2013}$ to illustrate how stakeholder positioning and power within an AISC may shape responses to an experienced harm.