The rapid ascent of large language models like ChatGPT, Gemini, and Claude has brought a long-simmering tension to the boiling point: the relationship between artificial intelligence and the decades-old laws that protect human creativity. At the heart of the controversy is the fact that these AI systems are trained on massive databases of published works, including hundreds of millions of books, online articles, and academic papers.
For many authors, this practice feels like a clear violation of their rights, as their intellectual property is used without consent to build tools that could eventually compete with them. However, recent legal developments suggest that the question of whether this training is actually illegal is far more complex than it appears on the surface.
The Anthropic Precedent: Lawful Training, Illegal Sourcing
One of the most significant moments in this legal saga arrived when Judge William Alsup ordered the AI firm Anthropic to pay a $1.5 billion copyright settlement to a group of writers. While the sheer size of the penalty seemed like a major victory for the creative community, the nuances of the ruling told a different story.
Judge Alsup did not rule that training an AI on copyrighted books was inherently illegal. In fact, he described the process of a large language model (LLM) ingesting trillions of words as being analogous to a human writer studying literature. He wrote that Anthropic’s models trained on these works not to replicate them, but to create something entirely different.
The massive fine was actually a penalty for how the data was acquired. Anthropic was found to have pirated the books from illegal online shadow libraries rather than obtaining them through authorized channels. Cathy Gellis, an attorney specializing in intellectual property and technology, suggests this ruling may actually favor AI companies in the long run. By framing the training process as reading or consuming a work rather than copying it, the court suggests that the act of learning from data is protected, even if the data collection method is not.
The Fair Use Dilemma and the 1976 Act
The primary hurdle for modern courts is that the governing legislation, the Copyright Act of 1976, was written nearly half a century ago. Judges today are forced to interpret guidelines from 1976 to address technologies that were then the province of science fiction.
The legal battleground usually centers on the concept of fair use. This is a legal carve-out that allows for the use of copyrighted material without explicit permission for purposes such as criticism, parody, education, and research. To determine if a use is fair, courts typically evaluate four factors:
- The purpose and character of the use.
- The nature of the copyrighted work.
- The amount and substantiality of the portion used.
- The effect of the use upon the potential market for the work.
Jason Henderson, Senior Attorney and Founder of the IP and Media Practice at JWL International, notes that the law is currently inconsistent. Because AI models are trained on such a vast scale, it is difficult for existing frameworks to provide clear answers, leading to what he describes as law that is all over the place.
Transformation vs. Competition: The Ross Intelligence Case
A critical factor in determining the legality of AI training is whether the new use is transformative. A transformative use adds something new, with a further purpose or different character, and does not substitute for the original use of the work.
The distinction between transformation and competition was highlighted in the case of Thomson Reuters v. Ross Intelligence. In that instance, the media company sued Ross Intelligence for using its content to build a competing AI-based legal platform. Judge Stephanos Bibas ruled against Ross Intelligence, stating that the use was not transformative because it shared the same purpose and character as the original Thomson Reuters content.
Henderson points out that if the goal of training an AI is to create a product that directly competes with the original property, courts are likely to view it unfavorably. While authors can argue that AI chatbots compete with them by generating synthetic books, this specific argument has struggled to find a definitive footing in court so far.
Distinguishing Training Data from AI Output
To understand the legal landscape, experts say it is vital to separate the act of training an AI from the act of copyrighting the content that the AI produces. These are two distinct legal issues with very different trajectories.
In the case of Thaler v. Perlmutter, the court established a clear boundary: works that are 100 percent AI-generated cannot be copyrighted. This ruling preserves the idea that copyright requires human authorship. However, it introduces a new set of practical challenges. As AI becomes more integrated into creative workflows, proving exactly what percentage of a work was assisted by AI becomes increasingly difficult.
Gellis compares the current situation to the use of basic digital tools. Most people feel comfortable with the idea that using spell check in a word processor does not strip a writer of their copyright. However, as AI tools move from simple corrections to generating entire paragraphs or chapters, the line between a tool and a creator becomes blurred.
The Problem the Industry Faces
The core problem for both AI companies and creators is the lack of a definitive solution. Most major AI firms are currently involved in pending litigation that will likely take years to resolve.
Initial rulings, such as the one involving Anthropic, are influential and help set the tone for the industry, but they are not the final word. Gellis notes that these early decisions could be undone if other courts reach different conclusions in the later stages of litigation. For now, AI companies must navigate a landscape of shifting legal interpretations, where a single ruling in one jurisdiction could force a massive change in how they handle data.
For authors and publishers, the goal remains the protection of the market for their work. If AI systems can ingest their life's work to create a competitor that renders the original work obsolete, the fundamental purpose of copyright law is called into question.
What Happens Next
The technology industry is currently in a period of high-stakes experimentation, both in the lab and in the courtroom. We are likely to see a series of conflicting rulings before any supreme legal consensus is reached or before Congress decides to update the Copyright Act for the 21st century.
Until then, the legality of training AI on copyrighted books will remain complicated. Companies will continue to argue that their training is a transformative act of learning, while creators will maintain that it is a systemic misappropriation of their intellectual labor.
The next few years of litigation will likely determine whether AI becomes a licensed partner to the creative arts or remains a controversial disruptor. As these cases move through the courts, the primary factor to watch will be how judges define the boundary between consuming information and copying it.
Filed under: AI, TechNews, Software, Startups, Copyright, LegalTech, IntellectualProperty