Reproduce “The Economics of AI Training Data: A Research Agenda” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hamidah Oderinwale, Anna Kazlauskas
The paper develops a research agenda for treating AI training data as an economic input alongside capital, labor, and algorithms. Combining conceptual analysis with documentation of 78 reported transactions, including 24 detailed deals from 2020–2025, it shows that data markets remain fragmented, largely bilateral, and highly dependent on quality, legal status, exclusivity, and domain; only 7 of the 24 detailed deals explicitly compensate original creators.
Key insights
- Data is nonrival in physical use but can exhibit emergent rivalry when duplication, aging, benchmark leakage, poisoning, or preference leakage reduces future usefulness.
- The verification paradox and legal uncertainty make data quality, provenance, copyright, consent, and permitted AI uses difficult to assess without risking copying, encouraging intermediaries and bespoke bilateral contracts.
- Data value depends on context, novelty, quality, legality, and interaction with existing data, compute, and labor; volume alone is not a reliable measure of economic contribution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.