Public data deals channel most value to aggregators while creators receive negligible royalties, exposing a structural data-value inequality; missing provenance, uneven bargaining power and static pricing threaten the sustainability of current ML pipelines and prompt a proposed EDVEX framework to redistribute benefits.

Main Finding

The paper argues the current machine-learning value chain systematically strips economic value from original data generators — an "economic data processing inequality" — because provenance is lost, bargaining is asymmetric, and price discovery is inefficient. By compiling and analyzing 73 publicly disclosed AI data deals, the authors show value concentrates with aggregators and model monetizers while creators rarely receive meaningful royalties. They propose an Equitable Data-Value Exchange (EDVEX) framework — combining task–data matching, auditable lineage, and utility-driven valuation — to create a more sustainable AI data marketplace.

Key Points

• Economic data processing inequality: technical signal improves along the ML pipeline (data → weights → synthetic outputs) but economic equity flowing to original creators shrinks.
• Three structural faults drive the inequality:
• Invisible provenance — lineage, consent, and license metadata are frequently lost as data is copied and repackaged, making auditing, attribution, and royalty routing infeasible.
• Asymmetric bargaining power — large aggregators and a handful of major buyers (e.g., OpenAI/Microsoft, Google) dominate negotiations; individual contributors rarely participate directly.
• Inefficient price discovery — many transactions are lump-sum buyouts with little or no revenue-sharing or usage-based pricing, disconnecting creators from downstream upside.
• Empirical signals from the 73-deal corpus:
  • Disclosed revenue sum (publicly reported portion): $677.3M.
  • 57 of 73 deals do not disclose revenue figures.
  • Only 6 of 73 deals mention any revenue-sharing with contributors; only one publicly reported an actual creator payout (~$2.5k).
  • Deals skew toward certain content types (news, images, academic, UGC) and a few major buyers.
• Consequences: continued exclusion of generators risks reducing data supply quality/diversity, concentrating market power, and undermining the feedback loop that fuels ML research and product development.

Data & Methods

• Data compilation: Assembled a dataset of 73 publicly disclosed AI data deals (detailed in the paper's appendix/Table 2). Sources were public press releases, reporting, and court filings where available.
• Descriptive analysis:
  • Coded deals by content category (e.g., news, images, academic, user-generated content), buyer identity, payment structure (disclosed amount vs undisclosed, recurring vs one-time), mention of generator splits, and litigation flags.
  • Produced summary statistics (e.g., counts by category and buyer; revenue disclosure rates) and visualizations (figures showing deal counts over time and revenue-share disclosure).
• Conceptual analysis:
  • Mapped how provenance loss, bargaining asymmetry, and static pricing interact along the ML value chain to generate the economic data processing inequality.
• Design proposal:
  • Developed EDVEX as a high-level framework and identified technical primitives and open research problems needed to operationalize it (task–data matching via sandbox utility estimation, lineage tracking with compact metadata, and valuation mechanisms that use estimated task-specific utility and enable revenue-sharing).

Implications for AI Economics

• Market structure and welfare:
  • Current deal practices push rents toward aggregators and a few platform buyers, increasing concentration (superstar effects) and potentially reducing overall market efficiency and innovation.
  • One-shot buyouts remove incentives for originators to maintain provenance or improve data quality, creating negative externalities for future model training and downstream trust.
• Data as an economic asset:
  • As datasets and derivatives become primary assets, absent provenance and fair pricing, markets will under-provide high-quality or consented data and over-rely on scraped or poorly governed sources.
• Research and infrastructure priorities:
  • Build scalable provenance tooling (lightweight, auditable metadata encodings) to enable attribution and royalties without excessive storage/compute burden.
  • Design and test sandbox-based task–data discovery and utility estimation methods that scale and generalize across modalities and tasks.
  • Develop market mechanisms (auctions, utility-linked pricing, revenue shares) and incentive-compatible allocation rules for dynamic, combinatorial data bundles.
• Policy and institutional responses:
  • Improved transparency standards for data deals, clearer obligations for platforms aggregating user content, and regulatory scrutiny of deceptive licensing practices could rebalance bargaining power.
• Risks and trade-offs:
  • Implementing EDVEX primitives must balance privacy, contributor incentives (avoiding premature disclosure of dataset value), and operational costs (compute and metadata storage).
  • There are open questions about how well sandbox utility estimates generalize and how to allocate value fairly across many small contributors when data value is combinatorial.

In sum, the paper diagnoses a structural market failure in AI data economies and lays out a research and engineering roadmap (EDVEX) to restore provenance, symmetric bargaining, and dynamic pricing so that data generators receive a meaningful share of the value they enable.