IP and Copyright

Introduction

Generative AI raises copyright questions from two directions at once. On the input side, most AI models were trained on enormous datasets scraped from the internet — books, news articles, images, music, and code — much of it copyrighted, collected without the creators' permission or payment. That practice is now working through courts and legislatures worldwide, with rulings that split even within the same year: one court has already ordered a $1.5 billion payout over pirated training data, while another dismissed a similar claim entirely. On the output side sits a separate question: when a prompt produces an image, an essay, or a song, who owns it — the person who typed the prompt, the company that built the model, or no one at all? U.S. law already has a clear, if narrow, answer to that question. This page walks through both debates, plus what happens when they collide: a model trained on someone else's copyrighted work generating something close enough to raise an infringement claim of its own.

The Debates

Whether training an AI model on copyrighted material without permission counts as fair use is being decided case by case, and the results don't line up neatly. Anthropic agreed to pay $1.5 billion — the largest copyright recovery in U.S. history — after it was found to have downloaded roughly 500,000 books from pirate sites like Library Genesis to train its models; the settlement was approved in September 2025, with a claims deadline of March 23, 2026 [1]. Meanwhile, The New York Times' case against OpenAI and Microsoft, over content scraped from the open web rather than pirate libraries, remains undecided: the Department of Justice has filed in the tech companies' favor, arguing that training delivers a "transformative public benefit," and a ruling on whether the case proceeds to trial is expected soon [2]. In the UK, Getty Images lost its core copyright claim against Stability AI when the court found the model doesn't retain a copy of the training images it learned from — though Getty won a narrower trademark claim over its watermark appearing in generated output [5]. A fourth case adds a wrinkle: Disney, Universal, and DreamWorks sued Midjourney in mid-2025, arguing its image generator doesn't just train on copyrighted characters but reproduces them almost exactly — simple prompts like "Yoda with a lightsaber" or "black armor with a lightsaber" returning near-perfect Yoda and Darth Vader images. That's a different legal theory than the training-data cases above: it's about the output looking too much like the original, not just the training process itself [6].

In 2025, the U.S. Copyright Office concluded that works generated entirely by AI in response to a prompt aren't copyrightable, because copyright protects human creativity and prompting alone isn't considered a creative act — even choosing your favorite among several AI outputs doesn't change that [3]. Work where a person meaningfully shapes the expressive result — editing, arranging, combining AI output with original material — can still be copyrighted, but only for the parts that reflect real human authorship. The courts agree: in Thaler v. Perlmutter, a computer scientist tried to register a fully AI-generated artwork with his AI system listed as the author. The D.C. Circuit rejected it in 2025, and the Supreme Court declined to hear an appeal in March 2026, closing the question for now [4]. The practical takeaway: a logo, paper, or image that's purely AI output — with no meaningful human editing — likely isn't protected by copyright at all, meaning technically anyone could reuse it, and claiming sole authorship over it raises its own honesty questions.

A related but distinct problem is AI that mimics a specific real person's voice, face, or artistic style rather than their literal copyrighted work. Federal law doesn't yet have a clear answer: the NO FAKES Act, which would create a federal right against unauthorized AI replicas of someone's voice or likeness, has been introduced four times since 2023 without passing, though its 2026 version — with new due-process protections and a carve-out for educational and library use — is being described as having its best chance yet [8]. In music, the major labels sued AI song generators Suno and Udio for over $500 million combined, arguing the tools could produce work convincingly "in the style of" real artists after training on unlicensed recordings. Universal Music Group and Warner Music Group have since settled and struck licensing deals with both companies — Udio pivoting to a closed platform built only on licensed music, Suno agreeing to license its training data and charge for downloads — while Sony's case remained unresolved as of late 2025 [7]. Licensing settlements, not court rulings, are currently doing more to define the rules here than legislation is.

Emerging Approaches

Rather than wait for lawsuits to resolve, many AI companies are now paying publishers directly for content. Over a dozen deals were struck through 2025 alone: Axios and OpenAI (OpenAI funded four new local newsrooms as part of the deal), the Associated Press and Google, Agence France-Presse and Mistral, The Guardian and OpenAI, and Meta licensing content from seven publishers including CNN, Fox News, and USA Today for its Llama models [10]. Some deals explicitly exclude the content from training data altogether, licensing it only for real-time summaries and links instead. This is fast becoming the default way large AI companies avoid litigation — though it mostly benefits publishers large enough to negotiate, not individual writers, photographers, or musicians.

Artists have adopted tools like Glaze and Nightshade, which subtly alter an image's pixels so AI models misread its content during training — Glaze alone has been downloaded roughly 7.5 million times [9]. But research published in 2025 introduced LightShed, a tool built by researchers at Cambridge, TU Darmstadt, and UT San Antonio that can detect and strip out exactly that kind of protection, transferring what it learns across different defense tools without needing to see them beforehand [9]. Even Glaze and Nightshade's own creators describe their tools as a deterrent and a statement, not a permanent shield — a pattern worth recognizing whenever a technical fix is offered as the final word against AI misuse.

The European Union's AI Act now requires companies that build general-purpose AI models to publish a public summary of the copyrighted content used to train them, using a standard template the European Commission released in 2025 — the first legal requirement of its kind anywhere [11]. No equivalent transparency law exists yet at the U.S. federal level, though similar disclosure requirements have been proposed in Congress and in several states. Whether transparency alone changes how AI companies source training data, or simply documents a practice that continues anyway, is still an open question.

Critical Questions to Consider

If you use AI to generate a logo, image, or piece of writing for a class project, a student organization, or your professional portfolio, is it copyrightable — and does that even matter to you? What would you need to do differently, according to the Copyright Office's standard, to make sure a piece is legally yours?

Anthropic paid $1.5 billion for training on pirated books, while The New York Times' case over content scraped from the open web remains undecided. What's the legal difference between those two situations? Does it feel like a moral difference to you too, or are they more alike than the different outcomes suggest?

Getty lost its UK case partly because the court found the AI model doesn't store a copy of the training images. Disney and Universal argue Midjourney can still reproduce Yoda and Darth Vader almost exactly from a simple prompt. Are these two lawsuits about the same underlying issue, or genuinely different ones? Which argument do you find more convincing, and why?

Pick one AI-publisher licensing deal — Axios and OpenAI, The Guardian and OpenAI, or another — and look up its actual terms. Who benefits most: the publisher, the AI company, readers, or none of the above? What would a similar deal look like for an individual creator instead of a large publisher?

Artists can use tools like Glaze to make their work harder for AI to train on, but researchers have already built a tool that can undo that protection. If a defense can eventually be defeated, is it still worth using? What would change your answer — and does the same logic apply to AI detection tools used in the classroom?

Curated Resources

These should be available in your college library.

Ivanova, I. (2024). Andy Warhol’s Factory made a fortune off other people’s work. Now desperate artists are accusing generative AI of supercharging that model. Fortune.Com, N.PAG. [BEGINNER]

Lindquist, D. (2023). In art, a question of ethics. Indianapolis Business Journal, 44(31), 4A-41A. [INTERMEDIATE] 

Lohrmann, D. (2023, March 26). Challenges With AI: Artistry, Copyrights and Fake News. Government Technology. [BEGINNER]

Singh, K. (2025, April 2). AI vs Art: Studio Ghibli Edition. Diplomat (Rozelle, N.S.W.). [BEGINNER]

Need Research Help?

Need help with your research?

Need help finding sources, checking a claim, or exploring a related angle? Your college librarians can help! Find your college specific libraries via the Maricopa Library page.

Instructor Resources — For Faculty

Assign small groups one real lawsuit each — Bartz v. Anthropic, New York Times v. OpenAI, Getty Images v. Stability AI, or Disney/Universal v. Midjourney. Each group researches its case's specific facts, then argues the side they're not naturally sympathetic to, in two minutes. The class votes on the most persuasive argument for each case, then compares it to the case's actual outcome — or predicts one, for cases still undecided.

Present students with three short scenarios: a fully AI-generated image, an AI-assisted essay with substantial human editing, and something in between. Using the Copyright Office's human-authorship standard, small groups decide which pieces would likely be copyrightable and defend their reasoning. Close by discussing where their own recent AI-assisted work would fall on that same scale.

References

[1] Copyright Alliance, "What to Know About the $1.5 Billion Bartz v. Anthropic Settlement," 2025; The Authors Guild, "Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement," Sept. 2025.

[2] Axios, "Historic NYT v. OpenAI copyright battle heats up," Sept. 8, 2026.

[3] U.S. Copyright Office, "Copyright and Artificial Intelligence, Part 2: Copyrightability," Jan. 29, 2025.

[4] U.S. Court of Appeals for the D.C. Circuit, Thaler v. Perlmutter, Mar. 2025; Holland & Knight, "The Final Word? Supreme Court Refuses to Hear Case on AI Authorship," Mar. 2026.

[5] Ropes & Gray, "Getty Image Loses Copyright Infringement Claim Against Stability AI in UK's First-of-its-Kind Ruling," Nov. 2025.

[6] Georgetown Law Institute for Technology Law & Policy, "Disney, NBC Universal, and DreamWorks File Major IP Lawsuit Against AI Image Generator Midjourney," Jun. 2025.

[7] Billboard, "What Do the Suno and Udio Licensing Deals Mean for the Future of AI Music?" Dec. 2025.

[8] Manatt, Phelps & Phillips, "Congress Reintroduces the NO FAKES Act: What's New in the 2026 Bill," May 2026.

[9] MIT Technology Review, "This tool strips away anti-AI protections from digital art," Jul. 10, 2025.

[10] Digiday, "A timeline of the major deals between publishers and AI tech companies in 2025."

[11] Pinsent Masons, "EU AI Act copyright template published," 2025.