Is AI Training IP Theft?

I have a bachelor’s degree in Music Composition from UCLA. When I was still trying to find my voice as a young composer, I did what countless students before me have done: I copied. I sat with scores by Berlioz, Brahms, Wagner, Stravinsky, Ravel, and Bernstein and transcribed passages verbatim. Then I created variations on the ones that spoke to me most. Then I began mixing those ideas together. Slowly, the results stopped sounding like Stravinsky and started sounding like me.

Artists in other fields describe the same journey. Painters copy the masters before they develop a personal style. VFX artists study and reverse-engineer existing sequences. Dancers absorb and recombine movement vocabularies. To kickstart creativity, we absorb and process influence. That is how the craft is learned.

AI models are doing this for us instantly. It used to take me a week to choose five passages from five great orchestral works and recombine them into something new. An AI model can do the equivalent five hundred times faster. The model is a calculator. Doing it by hand was long division on paper. The creative process itself has not changed; only the speed and scale of the absorption and recombination have.

Plagiarism, Slop, and Training Are Not the Same Thing

There is a great deal of confusion right now between plagiarism, “slop,” and training. AI models do make both plagiarism and low-effort output easier, just as the internet made the distribution of pornography easier. That was never a good reason to ban the internet or sue every ISP. The existence of a powerful tool that can be misused does not make the tool itself illegitimate.

Recent legal developments underscore the distinction. In July 2026 a federal judge approved Anthropic’s $1.5 billion settlement with authors—the largest copyright class-action settlement in history. The case centered on how Anthropic acquired books from pirate libraries such as LibGen and PiLiMi, not on whether training an AI model on copyrighted material is lawful. An earlier ruling in the same matter had already found that the training itself constituted fair use. The settlement required Anthropic to pay roughly $3,000 per claimed book and to delete the pirated files; it did not establish a general requirement that AI companies pay copyright holders simply for the act of training.

The Tool vs. the Output

AI companies have no obligation to pay copyright holders simply for training on publicly available or lawfully obtained content. Training does not create a derivative product that is sold as a substitute for the original works. It produces a general-purpose tool that creators can then use to generate potentially derivative works. Whether those works cross the line into plagiarism is a question about the final output and the user’s choices, not about the existence of the tool. If a human draws inspiration so closely that the result is substantially similar to a protected work—whether they used an AI assistant or not—they are the plagiarist. The AI is not primarily a plagiarism machine any more than the internet is primarily a pornography-delivery system.

The history of every major creative technology follows the same pattern. New tools lower the cost of imitation and of novelty at the same time. The responsible response has never been to outlaw the tool or to demand royalties for every act of learning. It has been to judge the finished work on its own merits and to enforce existing copyright law against actual copies and derivatives, not against the process of studying the culture that came before.