When Does AI Learning Become Copyright Infringement?

When does learning from someone else's work become copyright infringement? Should the answer be different when the learner is an artificial intelligence model?

AI companies have trained their models on enormous collections of books, articles, music, and other copyrighted material. Recent copyright lawsuits raise questions that go well beyond whether AI can legally learn from those works.

How were the materials obtained? Were they purchased, licensed, downloaded from publicly accessible websites, or acquired from pirated collections? Does the legality of obtaining a work determine whether it can legally be used for AI training?

In this solo episode, I examine the copyright litigation involving Anthropic and OpenAI, including revelations about pirated book collections, the distinction between acquiring copyrighted material and training on it, and why those activities may have different legal consequences.

I explore a question that receives less attention: What happens when AI models learn primarily from material that is popular or readily available, rather than material that is accurate, original, or valuable?

As an author whose own work has been implicated in the broader AI copyright controversy, I have a personal interest in these questions. I'm approaching them as an engineer trying to understand the problem before deciding what the solution should be. The challenge isn't determining whether AI training is legal or illegal. It's understanding where learning ends, where copying begins, and how we might protect creators without preventing technological progress.

Before we can solve the problem, we need to understand what problem we're trying to solve.

The episode isn't just about whether training AI on copyrighted books constitutes fair use. It's about how those books were acquired, what happened to the copies, and whether the resulting AI systems create different problems for authors and society.