Anthropic, the AI company behind the Claude models, was found to have illegally scraped and pirated the work of over 10 Stanford professors to train its systems. According to a ruling in the *Bartz v. Anthropic* lawsuit, the company obtained copyrighted texts from "shadow libraries" such as LibGen and PiLiMi. The court determined that downloading and storing these pirated files in a central library constituted a violation of copyright law. This finding contributed to a $1.5 billion settlement involving nearly 500,000 registered works, intended to compensate authors and publishers for the unauthorized use of their intellectual property.
This image illustrates the concept of a library of pirated works, which the court found to be a violation of copyright law.
Despite the ruling against piracy, Judge William Alsup distinguished between the illegal acquisition of data and the act of training the models themselves. The judge characterized the training of LLMs as "fair use," drawing a parallel to how human readers internalize themes and styles to eventually create new works. Under this interpretation, the AI is permitted to learn from the data to "turn a hard corner and create something different" rather than simply replicating or supplanting the original authors. This distinction creates a nuanced legal landscape where the method of data collection is strictly regulated, but the transformative use of that data in machine learning remains largely protected.
The settlement has not resolved all grievances, as some authors argue the compensation is insufficient and the precedent remains problematic. Professor Tobias Wolff, who opted out of the settlement, has characterized the use of his work without payment as a form of "slave labor" and expressed concern for the future of young writers. Meanwhile, legal scholars point to "transaction costs" as a primary hurdle, noting that it is practically impossible for AI platforms to obtain individual licenses for millions of unique works. As the Authors Guild pushes for national legislation to mandate licensing and output filters, the legal future of mass scraping remains a point of significant uncertainty for the creative industry.