AI Companies Propose Market Mechanism to Pay Creators
Harvard Business Review argues in a June 15, 2026 essay by E. Glen Weyl and Raul Castro Fernandez that AI companies already generate the two data sets needed to price creator content fairly during every training run: dataset composition (the mix of sources used) and scaling-law value signals (each source's contribution to model performance). The authors say the long-standing objection that data cannot be valued at scale no longer holds, and propose a market for future data access modeled on the music industry's collective management organizations (CMOs) rather than retrospective lawsuit payouts. Publishers, authors, and visual artists have argued their work was used without permission or payment, while AI firms counter that training on available data constitutes fair use. For practitioners, a functioning content market would reshape dataset-sourcing budgets, licensing risk, and procurement overhead.
The specific mechanism matters more than the general grievance: two named academics argue AI companies can price training data today, using data they already collect for other purposes, reframing the standoff between creators and AI labs as a solvable market-design problem rather than an intractable valuation one.
What happened
In a June 15, 2026 Harvard Business Review essay, E. Glen Weyl (RadicalxChange Foundation, Plurality Institute) and Raul Castro Fernandez (University of Chicago) argue that the dispute over AI training data, which they call one of the decade's defining economic conflicts, can be resolved through a forward-looking compensation market rather than retrospective litigation. Publishers, authors, and visual artists have argued their work was used without permission or payment; AI companies have countered that training on available data constitutes fair use and that pricing millions of individual creators' contributions is technically impossible.
Technical context
Weyl and Castro Fernandez say that objection no longer holds because AI companies already produce, as a byproduct of routine training, the two data sets needed to price content: dataset composition (the proportional mix of sources used) and scaling-law value signals (metrics revealing each source's marginal contribution to model performance). They propose adapting collective management organizations (CMOs), the licensing bodies used in the music industry, to distribute payments for future data access at scale.
For practitioners
If adopted, a functioning content market would change how ML teams budget for and source training data, shifting cost and risk from open-ended copyright exposure toward predictable licensing fees, while adding procurement overhead, provenance tracking, and compliance requirements to dataset curation pipelines.
What to watch
Track whether publishers, platforms, or AI labs begin adopting standardized provenance and metadata tagging, whether tooling emerges to extract auditable scaling-law value signals, and whether courts or regulators lean toward retrospective damages or forward licensing as the preferred remedy, since this proposal has not yet been adopted by any named company.
Key Points
- 1Harvard Business Review argues AI companies already generate the data needed to price creator content fairly during every model training run.
- 2The essay proposes a collective-management-organization model, borrowed from music licensing, for compensating creators through forward-looking data deals.
- 3Adoption would depend on standardized provenance and metadata tooling that does not yet exist across the AI training-data supply chain.
Scoring Rationale
A substantive, named-author proposal in a top business publication addressing a live and consequential AI-copyright dispute, relevant to how ML teams will budget for and license training data. Kept in the solid range rather than higher because it is a single-source opinion/analysis essay proposing an unadopted market mechanism, not a policy or technical change that has taken effect.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
