From Extraction to Exchange: A Data Streaming Framework for Equitable AI Content Economics
Primavera De Filippi¹, Sebnem Erener², Hannah London³
DOI : 10.65701/ofnfa9f1pb
Corresponding author:
pdefilippi@gmail.com
Abstract
Generative AI systems are trained on human content, mostly acquired without authorisation or compensation. This paper argues that existing legal and market responses reproduce existing power asymmetries, leaving content creators with a false choice between total exclusion and unconditional appropriation. Moving from training-time ingestion to inference-time retrieval opens a third pathway to the extent that it produces discrete content utilisation events that can be logged, attributed, and potentially compensated. We introduce “data streaming” as a specific configuration of inference-time retrieval enabling per-query compensation and data sovereignty. When compared with other emerging compensation architectures (per-query micropayments, per-crawl licensing, ad-revenue sharing), data streaming also presents some limitations, namely: its inapplicability to legacy content, unresolved attribution challenges, and the risk that micropayments normalise inadequate compensation. Insofar as market infrastructure alone cannot produce equitable outcomes, we argue that data streaming must operate within institutional conditions established through collective bargaining and regulatory intervention. In particular, because it falls outside the DSM Directive’s text-and-data-mining exceptions, we propose combining extended collective licensing with a statutory remuneration right.
