Work

Ghost algorithms? Subsidiary algorithmic work and negotiations with experts in the development of medical AI

Year: 2025
Type: conference-paper
Abstract: The advancement of predictive AI requires not only the development of algorithms by computer scientists but also the construction of the databases necessary for training those algorithms, commonly referred to as "groundtruths."In some applications, the process of creating groundtruths involves negotiations between computer scientists and domain experts.This article analyzes negotiations through two phases of the groundtruthing process involving expert collaboration that have received little scholarly attention.The first is the recruitment of experts (in our case, radiologists) by computer scientists; the second is the support provided to experts during the process of labelling the data needed to train the algorithms.Drawing on concepts from actor-network theory-specifically, problematization and passage points-we show that, throughout these negotiations, developers engage in what we term subsidiary algorithmic work: the design of nonfinalist algorithms that are simpler than the ultimate predictive models and are not intended for publication or commercialization, but rather serve to recruit or assist experts.These invisible computational actors-essential for securing expert participation but ultimately absent from formal outputs-are what we refer to as "ghost algorithms".
Language: English
Open Access status: green
OpenAlex ID: W7152350742
FWCI:
Cites: 49
Cited by:
Related to: 10