Externalities in AI Training

Large language models are trained on billions of documents, images, and code samples, much produced by individuals and organizations who were neither compensated for the use nor asked for consent. The AI companies and consumers benefit from the training-produced capabilities; the original creators bear a cost as their work builds competing capability potentially reducing their future output's market value. Under current property-rights assignment, this cost falls entirely on creators. The legal framework governing training-data use is unsettled, and transaction costs of enforcing copyright against AI training are prohibitive — the scale is measured in billions of documents, causal connections between individual documents and model outputs are diffuse and hard to establish, and legal frameworks were designed for reproduction, not statistical learning. The Coasian question is not whether creators deserve compensation but whether a different rights assignment would produce more efficient outcomes.