Indirect Crowdsourcing and the Hidden Flows of Data for AI

Abstract

AI systems depend on labelled datasets, yet producing high-quality labels remains costly and labour-intensive. An increasing share of “data for AI” is therefore generated by reusing the digital traces people leave as byproducts of everyday interaction. This work argues that this practice should be analysed systematically as a knowledge management (KM) phenomenon: it captures dispersed, situated know-how and codifies it into reusable knowledge assets. We formalise indirect crowdsourcing as the ex-post appropriation of human-generated digital traces into structured datasets or labels, characterised by unintentional participation, the absence of ex-ante task framing, and mediation or curation that enables reuse beyond the traces’ original purpose. We apply a unified analytic rubric—the Four Pillars of crowd, task, platform, and crowdsourcer—to three cases: Medical X, Pandemic Prediction, and reCAPTCHA. Using directed content analysis over publicly accessible artefacts and documentation, we reconstruct how traces move through a trace-to-asset lifecycle central to KM: production, selection and validation, codification and labelling, storage and access modalities, and downstream reuse. We further distinguish indirect crowdsourcing from adjacent phenomena relevant to KM and data work, including explicit crowdsourcing, human computation, and passive crowdsensing. Across cases, the analysis reveals recurring KM mechanisms (embedded versus ex-post mediation), differences in primary beneficiaries (platform owners, researchers, public agencies, etc.), and governance risks that undermine responsible knowledge reuse (opaque consent, licensing and provenance gaps, fairness concerns, and skewed value distribution). These findings highlight platforms as knowledge intermediaries that shape what can be captured, who can access it, and how value is allocated. Rather than offering a large-scale empirical evaluation of AI systems, the paper provides a conceptual and comparative analysis of how digital traces are converted into knowledge assets through indirect crowdsourcing. We contribute (i) an operational definition with observable implications for auditing indirect knowledge capture, (ii) a cross-case comparative rubric for evaluating trace-to-asset pipelines as KM processes, (iii) knowledge governance guidelines for responsible and equitable reuse, and (iv) a research agenda on boundary conditions, stewardship, and measurement for managing indirectly sourced knowledge at scale.