The Data-Production Dispositif

Abstract

Machine learning (ML) depends on data to train and verify models. Very often, organizations outsource processes related to data work (i.e., generating and annotating data and evaluating outputs) through business process outsourcing (BPO) companies and crowdsourcing platforms. This paper investigates outsourced ML data work in Latin America by studying three platforms in Venezuela and a BPO in Argentina. We lean on the Foucauldian notion of dispositif to define the data-production dispositif as an ensemble of discourses, actions, and objects strategically disposed to (re)produce power/knowledge relations in data and labor. Our dispositif analysis comprises the examination of 210 data work instruction documents, 55 interviews with data workers, managers, and requesters, and participant observation. Our findings show that discourses encoded in instructions reproduce and normalize the worldviews of requesters. Precarious working conditions and economic dependency alienate workers, making them obedient to instructions. Furthermore, discourses and social contexts materialize in artifacts, such as interfaces and performance metrics, limiting workers' agency and normalizing specific ways of interpreting data. We conclude by stressing the importance of counteracting the data-production dispositif by fighting alienation and precarization, and empowering data workers to become assets in the quest for high-quality data.

AI Empire: Unraveling the interlocking systems of oppression in generative AI's global order

“…The generative AI output companies such as Meta and OpenAI seek to optimize along with other data work only deepens the "precarization, alienation, and surveillance" of data workers (Miceli and Posada, 2022), as their labor, expertise, psyche, and infrastructures are exploited for this optimization. Such practices bear the hallmark of coloniality, but modernity is at play as well, with the West's "selfdelusion that it occupies an epistemic objectivity and thus may treat all other pieces of knowledge and ways of being as expendable peculiarities" (Mumford, 2022).…”

“…Through this process, data is turned into a commodity to be bought and sold in the big data marketplace of AI Empire as just another product in the hands of the data brokers responsible for "the Costcoization of data," where acquiring even the most personal data and analyzing it using advanced AI technologies is akin to going to a big-box store to purchase everything you need in one place (Lamdan, 2022). Extractivism does not stop at appropriating data; however, the exploitation of labor, natural resources and effects, and climate degradation are also an inextricable part of this logic (Crawford, 2021;Miceli and Posada, 2022;Sultana, 2022).…”

“…These processes reproduce old colonial/imperialist logic and place the burden of climate change disproportionately on the shoulders of the majority world as a form of climate coloniality (Sultana, 2022). Similarly, content moderation, data annotation, and the evaluation of AI output-processes that, as AI companies admit, are essential for the uncannily realistic responses generated by systems such as ChatGPT, 1 often function in settings reminiscent of colonial practices (Miceli and Posada, 2022). They are performed for poverty wages and at great psychological detriment by data workers in majority-world countries; in ChatGPT's case for instance, the "human guidance" part of the training and evaluation process was mainly performed by Kenya-based Sama and Majorel subcontractor workers for less than $2 per hour (Perrigo, 2023c).…”

“…Moreover, despite their seemingly "magical" performance, AI algorithms such as ChatGPT are far from fully automated. They rely heavily on the "ghost work" of human annotators and moderators, often in majorityworld regions (Gray and Suri, 2019;Miceli and Posada, 2022), as well as on massive media hype which tends to greatly overstate their actual capabilities or level of automation-a phenomenon known as "fauxtomation" or fake automation (Taylor, 2018). Another problematic trend is the perilous shift in focus away from the urgent real-world problems humanity is facing.…”

Investigating Labeler Bias in Face Annotation for Machine Learning

“…Geva et al [22] recommend that labelers for testing and training datasets be distinct groups since they found that subjective NLP labels produced by a group do not generalize well. Instructions for annotators have also been found to embed bias [47,42]. In the context of face annotation, Engelmann et al [15] argue that 'secondary' (i.e., subjective) characteristics may not be appropriate attributes for facial recognition systems to predict.…”