Source: Private Law Theory — Obligations, Property, Legal Theory

The paper distinguishes two distinct markets where artificial intelligence and copyright intersect: the market for access to training data and the market for reproduced outputs, and it argues these markets give rise to different welfare problems. It draws on the economics of information goods, technical literature on memorisation in large language models, and a comparative analysis of recent legal developments in Germany, the United Kingdom and the United States. Building on Gans (2024), the author proposes a capability-based levy that ties payments to rights holders to the measurable memorisation rate of deployed models so as to reflect potential harm in the market for reproduced outputs. The text emphasises that a capability-based approach focuses on observable model behaviour rather than solely on training-data provenance.

The paper identifies a distribution problem created by any such payment mechanism and highlights non-pecuniary harm to creators from unattributed use of their work. To address these concerns, the author argues for mandating source-attribution capability at the model-architecture level as a complement to monetary payments. The capability-based scheme is presented as having implementation advantages over training-data-based schemes, including reduced exposure to territorial arbitrage and stronger incentives for AI developers to reduce memorisation. A formal citation is given as Koboldt, Christian, Copyright and Artificial Intelligence: The Importance of Memorisation and Attribution (May 20, 2026).