AI Training Ruling Questions Authors Creative Value
The debate over whether AI companies can legally train their models on copyrighted books has produced a strange spectacle: lawyers arguing about the nature of reading, judges comparing machine learning to literary study, and authors watching their life’s work become fuel for the very tools that may soon replace them. But beneath this legal theater lies a more uncomfortable truth. The question was never really about copyright at all. It was about whether a human being’s unique contribution to culture can be replicated by a statistical process. And the courts, by focusing on the mechanics of copying rather than the nature of creativity, have already answered that question in ways that make the human author superfluous.
The Reading Problem
Judge William Alsup’s ruling against Anthropic in 2025 seemed like a victory for writers. The company was ordered to pay $1.5 billion to authors whose books were used to train its AI models. [1] The headlines wrote themselves: David defeats Goliath, creators strike back against the machines. But read the actual decision and something else emerges. Alsup did not find that Anthropic’s training was illegal. He found that the company had pirated books from shadow libraries, which is a different offense entirely. The training itself, he wrote, was lawful, comparing the way a language model ingests trillions of words to a writer’s study of literature.
This comparison deserves scrutiny because it smuggles in a dangerous equivalence. When a human writer reads a book, they absorb themes, techniques, and styles, then transform those influences through their own consciousness. The process involves judgment, taste, and the messy business of lived experience. An AI model does none of this. It processes text as data, extracting statistical patterns that allow it to predict which word comes next. The output may resemble human writing, but the process is fundamentally different. Yet the judge saw them as analogous, and that analogy has consequences far beyond the courtroom.
Cathy Gellis, an attorney specializing in intellectual property and technology, observed that the ruling treats AI training as “reading a copyrighted work as opposed to copying a copyrighted work.” [2] Copyright law, she notes, hinges on copying, not on using or experiencing a work. This distinction sounds reasonable until you consider what it means for the people who created those works. If training an AI on millions of books is legally equivalent to reading them, then the AI becomes a kind of super-reader, one that never forgets, never tires, and never needs to earn a living from what it has learned.
The Competition Question
The legal framework that judges are struggling to apply dates from 1976, a time when the internet did not exist and the idea of machine learning was confined to science fiction. Jason Henderson, Senior Attorney and Founder of the IP & Media Practice at JWL International, points out that the law has not caught up with the technology. [2] Courts are “all over the place” in their reasoning, he says, because they are trying to fit new questions into old categories. The outcome of these cases, he suggests, may depend less on legal principle than on whether the AI’s output competes with the original work.
This is where the Thomson Reuters case against Ross Intelligence becomes instructive. Ross built an AI-based legal research platform by copying content from Thomson Reuters, and the court found that this was not fair use because the new platform directly competed with the original. Judge Stephanos Bibas wrote that Ross’s use was not transformative because it did not serve a further purpose or different character than Thomson Reuters’s work. [3] The logic seems clear: if you use someone’s property to build a rival product, you have crossed a line. Yet this case concerns a direct commercial rival, not the broader question of whether a machine’s statistical learning diminishes the human act of creation.
But apply this logic to the broader AI landscape and the implications become unsettling. Authors could argue that chatbots compete with them by generating synthetic books from their works. Yet that argument has not prevailed in court, and the reasons are revealing. The courts have not yet decided that AI-generated books are close enough to human-written ones to constitute competition. The implicit assumption is that they are not, that there remains some quality in human writing that machines cannot replicate. This assumption may be comforting, but it is also fragile, and the legal system’s reliance on it suggests a deeper uncertainty about what exactly is being protected.
The Measurement Problem
Gellis suggests that the debate forces us to confront questions we have long avoided. If a work is 100 percent AI-generated, courts have ruled it cannot be copyrighted, as in the Thaler v. Perlmutter case. But this ruling raises a practical problem that no one has solved: how do we prove whether a work was generated by AI, and if it was, what percentage of it involved human input? The law demands clear categories, but the technology blurs every boundary. The deeper question is not legal but existential: if a machine can produce text that readers value, what remains of the author’s unique claim to cultural contribution?.
Consider the example Gellis offers. If you write a novel in Microsoft Word and run spell check, you feel comfortable saying that Word does not own your novel. The software is a tool, not a creator. But where is the line between tool and creator? If you use an AI to outline your novel, is it still yours? If you use it to write a draft that you then edit, what is the nature of your contribution? If you feed it your own unpublished manuscript and it produces something similar, who owns the result? These are not hypothetical questions. They are the everyday reality of writers trying to navigate a landscape where the tools of creation have become indistinguishable from the agents of creation.
The deeper issue is that copyright law was designed for a world where human creativity was the only source of new works. It assumed that the author was a person, that originality meant human originality, and that the market for creative works would be shaped by human choices. AI has shattered all three assumptions. The law now faces the impossible task of measuring something that resists measurement: the degree to which a machine can be considered an author, and the degree to which a human can be considered redundant.
The Final Consequence
The legal battles over AI training will continue for years, and the outcomes are far from certain. Gellis notes that the initial rulings are influential but could be undone by other courts deciding differently. The litigation is still in its early stages, and no definitive solution is in sight. But in the meantime, these decisions are shaping everything that happens in the industry, and it would be foolish for AI companies to ignore them.
Yet the most important consequence is one that no court has addressed. If AI models can learn from human works without infringing copyright, and if their outputs are not considered competitive with human creations, then the economic basis of authorship collapses. Why would anyone pay a human writer when a machine can produce text that is legally distinct, commercially viable, and infinitely scalable? The answer is that they would not, and the legal framework that has been built to protect authors may end up facilitating their obsolescence.
The judges who compare AI training to human reading may believe they are protecting creativity. In reality, they are describing a world where human creativity is no longer necessary. The author becomes a historical artifact, a relic of a time when the production of new ideas required a human mind. The law, by treating AI as just another reader, has already made the author superfluous. It simply has not caught up to the implication of its own reasoning. The copyright illusion is that the law protects creators. The reality is that it has become the instrument of their replacement. The question is not whether the law will adapt, but whether it can adapt without first admitting that the human author was never the point.
Sources
1. Anthropic