Participation versus scale: Tensions in the practical demands on participatory AI

Ongoing calls from academic and civil society groups and regulatory demands for the central role of affected communities in development, evaluation, and deployment of artificial intelligence systems have created the conditions for an incipient “participatory turn” in AI. This turn encompasses a wide number of approaches — from legal requirements for consultation with civil society groups and community input in impact assessments, to methods for inclusive data labeling and co-design. However, more work remains in adapting the methods of participation to the scale of commercial AI. In this paper, we highlight the tensions between the localized engagement of community-based participatory methods, and the globalized operation of commercial AI systems. Namely, the scales of commercial AI and participatory methods tend to differ along the fault lines of (1) centralized to distributed development; (2) calculable to self-identified publics; and (3) instrumental to intrinsic perceptions of the value of public input. However, a close look at these differences in scale demonstrates that these tensions are not irresolvable but contingent. We note that beyond its reference to the size of any given system, scale serves as a measure of the infrastructural investments needed to extend a system across contexts. To scale for a more participatory AI, we argue that these same tensions become opportunities for intervention by offering case studies that illustrate how infrastructural investments have supported participation in AI design and governance. Just as scaling commercial AI has required significant investments, we argue that scaling participation accordingly will require the creation of infrastructure dedicated to the practical dimension of achieving the participatory tradition’s commitment to shifting power.