The AI Copyright Wars
Does training AI models on copyrighted material constitute fair use, and who should capture the economic value created by generative AI built on others' creative work?
Canonical Synthesis
Author: terry-tang | Last updated: 2026-03-08
The AI copyright wars are the legal dimension of a fundamental question about the generative AI era: who owns the value that AI creates from human creativity? By 2026, this question remains unresolved by courts but has already been substantially resolved by economics — AI companies trained on copyrighted material first and are now litigating from a position of accomplished fact.
The litigation landscape is extensive. The New York Times v. OpenAI (December 2023) is the flagship case, combining a prestigious plaintiff, strong evidence of near-verbatim reproduction, and damages potentially reaching billions of dollars. Parallel suits from the Authors Guild, visual artists, Getty Images, music labels, and individual creators have opened multiple fronts. The EU AI Act's requirement that AI companies respect European copyright law adds a regulatory dimension.
The core legal question is deceptively simple: is training an AI model on copyrighted text, images, or audio a "fair use" (under US law) or equivalent permissible use? AI companies argue yes — that training is "transformative" because the model learns patterns rather than copying specific works, analogous to a human reading books and learning to write. Plaintiffs argue no — that training involves making copies of copyrighted works without permission, and that AI outputs directly compete with the originals.
The practical situation is murkier. AI companies have simultaneously argued that training on copyrighted data is fair use while signing licensing deals with some publishers — a position that plaintiffs have characterized as contradictory. If the data has no compensable value, why pay for it? If it does have value, why should only some creators be compensated?
Meanwhile, the economic displacement that copyright law is meant to address is already occurring. AI-generated text competes with freelance writing. AI image generators compete with illustrators and stock photographers. Google's AI Overviews reduce clicks to publisher websites. AI coding tools reduce demand for some programming work. Whether or not courts eventually rule on the training question, the economic impact on creators is real and accelerating.
The Arc
2023: The Opening Salvo. The legal campaign began in earnest with suits by artists against Stability AI and Midjourney (January 2023), escalated through the Authors Guild class action against OpenAI (September 2023), and culminated in the New York Times' landmark lawsuit (December 2023). These early suits established the legal theories and framed the debate.
2024: Proliferation and Complexity. Copyright litigation proliferated across media types — text, images, music, code, and voice. The Scarlett Johansson incident (May 2024) added a celebrity dimension and raised voice-likeness rights. Google's AI Overviews launch (May 2024) demonstrated the economic displacement in real time. The EU AI Act (March 2024) required transparency about training data and compliance with EU copyright. Courts began ruling on procedural issues but no substantive fair use decisions emerged.
2025-2026: Waiting for Precedent. As of early 2026, no court has issued a definitive ruling on the core fair use question for AI training. The cases are grinding through the legal system, with potentially years of litigation ahead. Meanwhile, AI companies continue to train on internet data, and licensing deals with willing publishers create a patchwork of compensation that leaves most creators uncompensated.
Interpretations
Fair-use reading
Training AI models on publicly available content is transformative use that benefits society. Just as humans learn from reading copyrighted books without paying royalties on their education, AI models learn patterns from data without copying specific works. Requiring permission or payment for training data would make AI development prohibitively expensive and concentrate it among companies wealthy enough to negotiate universal licenses — the opposite of democratization.
Proponents: AI companies, some legal scholars, tech industry advocates.
Creator-compensation reading
Training on copyrighted material without permission or compensation is theft at scale. AI companies have built multibillion-dollar products on the unpaid creative labor of millions of writers, artists, musicians, and other creators. The resulting AI systems directly compete with those creators, depressing wages and eliminating jobs. A system of mandatory licensing or revenue sharing is necessary to ensure creators benefit from the value their work generates.
Proponents: Authors Guild, media companies, visual artists, musicians, copyright maximalists.
Structural-reform reading
The copyright frame is too narrow. The real issue is the power asymmetry between AI companies and individual creators. Even if courts rule in creators' favor, the practical challenges of enforcing copyright against AI training are enormous. What's needed is not just legal rights but structural reforms: data dividends, collective bargaining for creators, public-interest AI models trained on licensed data, and regulatory frameworks that address the economic displacement directly rather than through copyright doctrine alone.
Proponents: Digital rights organizations, labor advocates, some legal scholars, public-interest technologists.
Open Questions
- Will the NYT v. OpenAI case produce a clear fair use ruling, or will it settle, leaving the legal question unresolved?
- Can a licensing-based approach to AI training data work at scale, or is the transaction cost of negotiating with millions of creators prohibitive?
- Does the EU AI Act's copyright compliance requirement create a practical difference, or will enforcement prove impossible?
- How should voice, likeness, and style be protected when AI can replicate them without directly copying any specific work?
- Is there a credible path to compensating creators that doesn't simply entrench the negotiating power of the largest publishers and labels?