Mapping Public Data Deals – Add Your Own (research.brickroad.network)

Ruoxi Jia et al. argue that a sustainable AI economy depends on clear, fair “data deals” that serve the people who generate training data. In a NeurIPS position paper they map and analyze publicly known data deals—highlighting inconsistent licensing, missing provenance, weak attribution, and few practical compensation pathways for creators—and introduce a crowdsourced registry to document and crowd-annotate deal terms. The paper frames these gaps as structural risks: opaque datasets undermine consent, auditing, and long-term incentives for content creators, while making model builders legally and ethically exposed. Technically, the authors outline pragmatic building blocks for working data deals: standardized rights/usage metadata, provenance tracking and cryptographic signatures to prove origin, machine-readable license templates, valuation heuristics, and privacy-preserving sharing (e.g., differential privacy and selective disclosure). They also discuss enforceability and auditing (verifiable compute logs, dataset manifests) and mechanisms for automated revenue- or attribution-sharing. The proposal is lightweight and actionable—aimed at dataset curators, model developers, and policymakers—to improve transparency, enable reproducible audits, and align incentives so creators get recognized and compensated without stalling model innovation.