Fair use decisions in copyright law rely on four factors: purpose of use, nature of work, amount taken, and market effect.
Can an AI legally learn from copyrighted books? Recently, a federal court tackled this issue in a case involving Anthropic's AI training process. And their answer is yes.
Copyright encourages creativity by granting exclusive rights, but these rights must balance incentive and public benefits. From the Stanford Library, U.S. judges assess fair use through four key factors:
The Purpose and Character of Your Use
How is the work being used? Transformative use, where the new work changes the original purpose or adds new insights, typically favors fair use. For instance, parodies or educational critiques are transformative.
The Nature of the Copyrighted Work
Is the original work factual or creative? Factual content, such as biographies or news articles, usually receives less protection, making fair use more likely. Creative works, such as novels and movies, receive stronger protection.
The Amount and Substantiality of the Portion Taken
How much content was used, and how important is that content? Even a small portion that's central or crucial, such as a song's iconic chorus, might not qualify as fair. On the other hand, using a larger amount that's necessary for commentary or criticism can still be fair.
The Effect of the Use Upon the Potential Market
Does the new work negatively affect the original's market value? If the secondary work directly competes with or replaces the original, judges usually rule against fair use. But if it targets a different market or enhances interest in the original, fair use is more likely.
In evaluating the fair use of AI, judges emphasized the transformative nature of training.
AI doesn't reproduce books. Instead, it learns to generate original content, similar to human education. Judges also acknowledged the necessity of using vast amounts of content for training, distinguishing this from direct copying.