The debate over artificial intelligence and copyright law is heating up, particularly regarding large language models (LLMs) trained on copyrighted material, including song lyrics. A central argument from copyright holders is that LLMs, by being exposed to copyrighted texts, are engaged in a form of unauthorized reproduction. However, I argue that training AI models on copyrighted material should not be considered illegal. In fact, it aligns with the way humans learn, reference, and incorporate prior knowledge into creative works.
Training vs. Copying: A Crucial Distinction
Copyright law is built around the principle that copying and redistributing protected works without permission is illegal. However, training is fundamentally different from copying. When an LLM is trained on a dataset that includes copyrighted lyrics, it does not store verbatim copies of those lyrics in its memory. Instead, it analyzes patterns, structures, and linguistic features—much like a songwriter who absorbs influences from the music they grew up with. If this type of learning were considered infringement, it would also be illegal for a human to memorize lyrics or learn songwriting techniques from their favorite artists.
The Human Parallel: How We Learn and Create
Consider how human musicians and lyricists develop their craft. Every songwriter is shaped by the music they have heard before. Artists reference previous works in their own songs through homages, allusions, and even direct lyrical references. If a songwriter internalizes the themes and patterns of Bob Dylan or Joni Mitchell and later writes lyrics that reflect similar ideas, we do not consider that illegal—because it is not a direct copy but rather an influence. AI models operate in much the same way, yet they are being held to an entirely different standard.
Fair Use and the Role of AI in Music Creation
Under U.S. copyright law, fair use allows for the use of copyrighted material in transformative ways, such as research, commentary, and education. Training an AI model falls within these categories. The purpose of AI training is not to regurgitate copyrighted material but to learn general language and structure. This is akin to how students study literature, songwriters analyze lyrics, and critics reference musical works in their reviews.
The Slippery Slope of Banning AI Training on Copyrighted Works
If the argument that training equals infringement were upheld, it could set a dangerous precedent. Would a university course that teaches music students by analyzing classic songs be considered illegal? Would music producers who reference old recordings for inspiration be guilty of copyright violations? These restrictions could stifle creativity rather than protect it.
AI-generated content that directly reproduces copyrighted lyrics should, of course, be subject to copyright laws—just as plagiarized works by humans are. But the process of learning and absorbing from copyrighted works should not be equated with copying. To claim otherwise is to misunderstand both the nature of AI and the essence of creativity itself.
Conclusion: Training on Copyrighted Lyrics Should Be Fair Use
The real issue is not whether AI models are exposed to copyrighted materials but whether they directly reproduce them in a way that competes with or replaces the original work. Training AI on copyrighted lyrics should be seen as part of the natural evolution of artistic learning—one that mirrors the way human creators refine their craft. If we outlaw this, we risk drawing an arbitrary line between human and machine creativity that does more harm than good.
Instead of restricting AI training, the conversation should focus on how AI-generated content is used. The key legal distinction must remain between learning from versus copying copyrighted works. Anything else threatens not only AI development but the very nature of creative inspiration itself.
