From Prediction Markets To Capability Markets: Epistemic Extractivism In Generative AI

Abstract

Some Generative AI (GenAI) platforms are introducing advertising into their conversational interfaces, others have committed not to. We use this divergence as a starting point, and argue that it obscures a deeper convergence: all major GenAI platform operators instrument the interfaces through which users interact with GenAI models, and all are positioned to develop and monetise AI capability products from the records this instrumentation produces. Drawing on analysis of the infrastructure that standardises and structures the recording of GenAI interactions, alongside industry documentation, research, and reporting, this paper surfaces interaction telemetry data as a distinct data object in the political economy of GenAI: a record of the cognitive processes materialised at the GenAI interface. We identify two forms of extraction co-constituted with this data object. Behavioural extraction, continuous with surveillance capitalism, captures user states for the prediction and shaping of user behaviour. Cognitive extraction, in contrast, captures teleological labour -- the labour through which users cajole GenAI platforms into doing what they want -- and refines this into AI capability products. In this context we draw on Grosfoguel's account of the appropriation of knowledge from marginalised communities to argue that cognitive extraction is an example of epistemic extractivism, and anticipate that its impacts will similarly be patterned by colonial legacies. Cognitive extraction positions human reasoning as a labour source to be extracted, and labour markets as a stock to be depleted. We conclude by outlining research directions for the study of epistemic extractivism in GenAI.