A Sustainable AI Economy Needs Data Deals That Work for Generators

AI summary

Overview

Research area: Machine learning economics, data governance, and AI policy — specifically the market structures that move data from its creators to model developers.

Technical level: Beginner-Friendly. This is a position paper. It presents no new model, no training experiments, and no benchmark results; its evidence is a qualitative and descriptive analysis of publicly disclosed business deals, supplemented by a proposed conceptual framework.

Scope (one sentence): The paper argues that the machine learning value chain is structurally unsustainable because value systematically flows away from data generators, and it proposes a three-part framework — the Equitable Data-Value Exchange (EDVEX) — to repair that flow. Authors are affiliated with Virginia Tech, Brickroad, Princeton University, and UC Berkeley (arXiv:2601.09966v1 [cs.LG], 15 Jan 2026).

What This Paper Is About

Machine learning runs on a chain that starts with raw data and ends with monetizable products, but the money does not travel back down that chain. The authors call this an "economic data processing inequality": at each stage — from inputs, to model weights, to synthetic outputs — the technical signal is refined while economic equity is stripped from the people who produced the original data.

The paper's goal is to document that inequality with evidence from real deals, name the three mechanisms that cause it, and sketch a market design that would let data generators bargain and be paid fairly.

Key Contributions

1. An empirical audit of 73 publicly disclosed data deals. The authors compile and analyze a collection of 73 deals (full list in Appendix A) and use it to show how missing lineage, weak bargaining power, and one-shot pricing reinforce each other into a cycle that concentrates capital and cuts generators out.

2. A diagnosis of three structural faults. The paper identifies invisible provenance, asymmetric bargaining power, and inefficient price discovery as the operational machinery of the inequality, and argues they interact multiplicatively rather than independently.

3. The EDVEX Framework. A blueprint with three layers — task-data matching, lineage tracking with auditable provenance, and utility-driven valuation — designed to patch the pipeline so that it provides bargaining symmetry, provenance, and efficient pricing simultaneously.

4. An open-problem agenda plus engagement with counterarguments. The paper lists concrete research problems for each EDVEX layer, compares its position to related work (data dignity, data cooperatives, online marketplaces, regulation), and answers opposing views including synthetic data and "data-for-service" models.

Main Findings

• Creator royalties round to zero. Of the $677.3m in reported revenue across the compiled deals, creator royalties round to almost zero. The authors describe this contrast as putting the scale and distribution of value in "crass contrast."

• Most deal terms are dark. 57 of the 73 found deals do not disclose any revenue publicly. Figure 3 shows 15 deals where public sources quoted a revenue volume (solid segment) and 52 additional deals without revenue share information (dashed segment). Table 1 lists 16 deals with amount disclosed, 57 undisclosed, 3 recurring, 6 with a generator split, and 4 under litigation.

• Revenue sharing is rare and tiny when known. Six of the 73 deals indicate revenue sharing with generators, and only one has public information on the actual revenue: $2.5k.

• Buyer concentration is severe. More than 40% of known transactions were conducted by OpenAI/Microsoft, Google, or Perplexity on the buying side. Table 1's top buyers are OpenAI (24 deals), Undisclosed (8), Google (6), and Perplexity (3).

• Deal types skew toward a few categories. Table 1 reports News 26, Images 16, Academic 15, and UGC 14 among the top types.

• Provenance failures are widespread and old. An analysis of over 1800 datasets found that more than 70% have license omissions, and datasets with licenses have error rates of over 50%. As early as 2017, the DeepMind–Royal Free case saw the UK Information Commissioner rule that data generators "were not adequately informed that their data would be used" in deep learning products.

• Aggregators capture licensing revenue without passing it on. Reddit banked $203 million in data licenses until early 2024 yet channeled $0 to the volunteers who wrote the content, and separately receives $60 million per annum from Google.

• Payments are typically one-shot buy-outs. Associated Press agreed a two-year, flat-fee deal (amount undisclosed), and News Corp settled for roughly $250 million across five years. To the extent of public disclosure, these contracts do not appear to include royalty escalators for generators tied to usage, retraining, or downstream revenue.

• Some deals were struck under legal pressure rather than negotiation. Several deals in the appendix are marked "L" for litigation, signaling they were hedged under the prospect of court action. The paper cites Getty Images v. Stability AI, NYT v. OpenAI, and News/Media Alliance v. Cohere as ongoing cases.

• The cost of getting this wrong can dwarf the data market. Meta's settlements — $1.4bn in Texas for biometric data misuse and $725m for the Cambridge Analytica case — dwarf the estimated $677m value of data deals in Figure 3.

Methodology in Plain English

The authors take a position rather than running experiments. They assemble a table of 73 data deals that have been publicly disclosed, drawn from public listings and filings, and record for each deal the data receiver, the data aggregator, the reference, the date, the type (News, Images, Academic, UGC), the disclosed dollar value, and codes such as "L" for litigation, "C" for confidentiality, "S" for settlement, "R" for recurring, and "U" for undisclosed.

They then count and categorize these deals — how many disclose money, how many share revenue with generators, who the buyers are — and read those counts as symptoms of three deeper mechanisms. Those mechanisms are then decomposed in a section each, drawing on additional public cases (Reddit, Stack Overflow, Shutterstock, news publishers, and litigation) to illustrate how they cascade through the value chain.

From that diagnosis they derive the EDVEX Framework on paper: three layers that would, in principle, match data to tasks, track lineage, and price data by measured utility. Alongside the design they write out open research problems and compare their proposal to related and opposing positions. No implementation or pilot is presented.

Why This Matters

Impact on research. The paper argues that excluding data generators shrinks the supply of high-quality, diverse data, which would leave researchers, startups, and large labs facing fewer and less representative datasets — the fuel current learning algorithms require. It also frames provenance loss as a problem that affects not just training but inference on model weights, since models can regurgitate images and text the provider may not have a license for.

Real-world applications:

• Licensing negotiations between publishers and AI developers, where utility-based pricing could replace flat, undisclosed fees.
• Royalty routing and attribution systems for user-generated content platforms, stock libraries, and code repositories.
• Data unions and cooperatives that form around specific machine learning tasks rather than fixed memberships, as in the healthcare cooperative MIDATA cited in the paper.
• Compliance tooling for regulations such as GDPR and the EU Data Act, which the authors argue EDVEX could help operationalize.

Industry relevance. Model monetizers are explicitly included as beneficiaries, not targets: the paper argues that buyers who negotiate with aggregators rather than generators directly in an open market are themselves getting a worse deal than an open market would give them. The authors also point out that adopting such a framework could reduce legal and financial exposure, given that Meta's settlements dwarf the total disclosed value of the data deal market in their sample. New market opportunities are identified for platform developers and infrastructure providers building discovery engines, secure sandboxes, lineage trackers, and valuation tools.

Future Directions

1. Scalable provenance and lineage tooling. What specific information should be logged, how granular should the metadata be, how can encoding schemes represent combinations of sources from potentially millions of data creators, and how can metadata size be traded off against tracking accuracy — all without manual effort from practitioners who use frameworks like PyTorch?

2. Reliable pre-acquisition utility estimation. How can a "sandbox-first evaluation protocol" using representative subsamples and small proxy models scale to potentially millions of datasets and thousands of tasks, and how well do the scaling-law-based extrapolations generalize across data modalities (tabular, time-series, graph), architectures, and complex tasks such as reinforcement learning?

3. Fair revenue sharing inside data unions. How should proceeds be split among contributors — the paper points to cooperative game theory concepts such as Shapley values based on marginal contribution — and how can valuation mechanisms be designed to mitigate "gaming" while rewarding genuine data quality, fairness, and robustness?

4. Price erosion for substitutable data. For data that is genuinely highly commoditized and where individual contributions offer little unique marginal utility even within optimal bundles, the risk of price depression remains an open area the authors explicitly flag as unresolved. A related open question is how synthetic data and iterative training loops should be treated, given that the authors argue realistic, long-tail, and physical-world data cannot be fully replaced.

Target Audience

This paper is most useful to people who sit between machine learning and economics: AI policy researchers, data governance and compliance teams at AI companies, platform and marketplace product leaders, data licensing lawyers, and data cooperative organizers. It is also valuable for machine learning researchers who want a concise, evidence-backed statement of where their training data comes from and who is paid for it.

Because it is a position paper with no experiments, it suits readers looking for a framing and an agenda — including undergraduate or graduate seminar discussions — rather than readers seeking an algorithmic technique. The authors are candid about limits: the evidence base relies on publicly disclosed deals and public filings, many transactions are private or under NDA, the dataset likely undercounts and skews toward larger, English-language, and U.S.-centric agreements, and completeness cannot be assumed.

Authors’ abstract

We argue that the machine learning value chain is structurally unsustainable due to an economic data processing inequality: each state in the data cycle from inputs to model weights to synthetic outputs refines technical signal but strips economic equity from data generators. We show, by analyzing seventy-three public data deals, that the majority of value accrues to aggregators, with documented creator royalties rounding to zero and widespread opacity of deal terms. This is not just an economic welfare concern: as data and its derivatives become economic assets, the feedback loop that sustains current learning algorithms is at risk. We identify three structural faults - missing provenance, asymmetric bargaining power, and non-dynamic pricing - as the operational machinery of this inequality. In our analysis, we trace these problems along the machine learning value chain and propose an Equitable Data-Value Exchange (EDVEX) Framework to enable a minimal market that benefits all participants. Finally, we outline research directions where our community can make concrete contributions to data deals and contextualize our position with related and orthogonal viewpoints.