Today, General Counsel and Senior Fellow Tim Hwang, with help from Research Fellow Sam Roland, filed a proposed amicus brief in support of OpenAI in In re OpenAI Copyright Infringement Litigation.
Introduction
General AI capabilities already serve important national security purposes. Over the last several years, the Government has acquired and adapted commercial frontier models for intelligence analysis, cyber defense, operational planning, and decision support. Both branches of government understand the importance of this new technological paradigm: Congress has supported it by tying defense AI to training and data and the executive by proclaiming rapid access to current models as one of its top national security priorities.
Copyright protects authors for a simple reason: their work advances knowledge and culture. Fair use serves the same end. It preserves room for new uses that benefit the public without displacing the expression copyright protects. But factor four of the fair use test is not limited to an accounting of the copyright holder’s potential lost revenue. Under Google LLC v. Oracle Am., Inc., courts may also consider the public benefits that copying enables, particularly where the use puts copyrighted works toward new productive ends. To the extent AI training enables the creation of new knowledge, expression, software, and other capabilities, those benefits bear on factor four’s assessment of the use’s market effects. That reasoning supports a novel application here: considering the national-security value of capabilities that enable new analysis, software, and other productive work.
The link between copying a newspaper article and developing national-security capabilities lies in the training process. During pretraining, models identify shared patterns across a broad corpus of information. Learning those patterns helps develop language, synthesis, translation, coding, and reasoning capabilities.
Requiring permission for the challenged training could impair those benefits if it materially restricts access to useful training data. Removing useful sources from training data can reduce capabilities used by downstream systems, which could include the national security systems the Government acquires and adapts. Under Oracle, those public benefits belong in the factor-four balance alongside cognizable market harm.
