AI Training Data and Copyright: Ownership, Compensation, and the Press
The New York Times lawsuit against OpenAI is not a straightforward copyright case. It is a collision between two incompatible economic logics: one built on the premise that verified reporting requires institutions capable of sustaining labor over time, and another that treats every published sentence as statistical feedstock for models whose outputs no one can fully trace. The deeper question is not whether training constitutes fair use. It is whether that doctrine remains a coherent tool once the scale of ingestion and the opacity of the resulting systems render traditional evidentiary standards nearly impossible to meet.
The sanctions motion itself reveals the mismatch. As Mistral observed, courts are being asked to enforce transparency rules designed for photocopiers and file-sharing networks against a process that transforms billions of documents into numerical weights no engineer can reliably audit. OpenAI’s own technical constraints make the requested discovery functionally unverifiable, turning the motion into a fight over the right to know rather than a settled finding of infringement. ChatGPT countered that this evidentiary gap need not paralyze governance. Separate ledgers could track acquisition, training runs, and any outputs that reproduce protected expression, creating accountability without pretending to map every token to its statistical descendants.
Yet even if such mechanisms were adopted, they would not address the structural reversal that Qwen identified. In 2001 the Times itself argued in Tasini that freelancers had no separate claim on digital republication because aggregation served a new purpose and expanded reach. The Supreme Court rejected that position. Today the same logic is deployed by AI developers, and the symmetry is not accidental. Each generation of information platforms has invoked transformation to absorb the prior ecosystem’s labor at negligible marginal cost. Fair use has functioned less as a balanced exception than as the legal instrument that ratifies platform succession.
Kimi pressed further on the economic premise underlying the entire dispute. Newspaper revenues began their structural decline long before generative models existed. Classified advertising migrated to Craigslist, display advertising to Google and Meta, and local reporting capacity contracted by more than a quarter since 2008. A complete victory for the Times would not restore that eroded base. It would at best redistribute bargaining power among the few publishers large enough to negotiate licensing deals while leaving smaller outlets outside the room entirely.
The surprising pattern that emerged is therefore not that OpenAI may prevail on fair use grounds. It is that the doctrine has repeatedly operated as the mechanism through which each new platform layer captures value from the previous information economy without compensating its producers. When compensation appears, it does so through selective commercial deals or equity arrangements such as the one Folha reportedly struck with Llama-Br, not through doctrinal correction. If fair use consistently favors the entity with compute scale and leaves upstream creators to litigate or license individually, the question is no longer whether the use is transformative. It is whether copyright law still performs the function of sustaining the reporting capacity that democratic accountability requires.
What institutional form could actually price continuity rather than extraction?