A landmark Supreme Court refusal, an Italian first-in-class statute, and billions in settlement money. The fight over who owns AI-generated work is rewriting intellectual property law in real time.
The Billion-Dollar Question Nobody Predicted
In February 2023, a graphic novel called Zarya of the Dawn forced the United States Copyright Office into unfamiliar territory. The book’s author, Kris Kashtanova, had used Midjourney to generate the illustrations. The Copyright Office ruled that the text and the arrangement of images could be copyrighted, but the individual AI-generated images could not. It was a split decision that satisfied nobody and clarified almost nothing.
Three years later, the confusion has only deepened. On March 2, 2026, the U.S. Supreme Court declined to hear Thaler v. Perlmutter, the case that asked a seemingly simple question: can an artificial intelligence system be listed as the author of a creative work? The answer, for now, is no. But the reasoning behind that answer, and the massive gaps it leaves open, have turned intellectual property law into a field where almost every assumption is up for grabs.
What makes this so consequential is the scale. Generative AI tools now produce millions of images, articles, code snippets, and musical compositions every day. If none of that output is copyrightable, it sits in a legal no-man’s-land where anyone can use it and nobody can protect it. If it is copyrightable, we need to figure out who the owner is. Neither outcome is straightforward.
What Courts Have Actually Decided
The legal landscape is defined by a handful of decisions that each draw a different line in the sand.
The Thaler saga started in 2018 when Stephen Thaler filed copyright registration for a visual artwork he said was autonomously created by his AI system, DABUS. The Copyright Office rejected it. Thaler sued. A federal district court ruled against him. The D.C. Circuit Court of Appeals affirmed, holding that human authorship is a bedrock requirement for copyright registration. When the Supreme Court refused certiorari in March 2026, it effectively sealed this principle into American law for the foreseeable future.
But the Thaler case dealt with a narrow scenario: AI acting entirely on its own, with no human creative direction. The harder question is what happens when a person uses AI as a tool. The Zarya of the Dawn registration established that human-authored elements within an AI-assisted work can receive protection, but the boundary between “human-directed” and “AI-generated” remains blurry.
Meanwhile, the training data cases drew even more attention. In Thomson Reuters v. ROSS Intelligence, a federal court ruled in February 2025 that legal headnotes were protectable and that copying them to train an AI model was not fair use. Then, in June 2025, two landmark rulings reshaped the debate. In Bartz v. Anthropic, Judge William Alsup drew a sharp distinction: training AI on legally acquired books qualifies as fair use, much like a human reading widely to learn to write. But downloading pirated copies from shadow libraries does not. Anthropic settled the resulting class action for $1.5 billion in August 2025, the largest AI-related legal settlement to date.
How Other Countries Are Drawing Different Lines
The United States is not the only jurisdiction grappling with this. And the answers are diverging, which creates a problem for any company that operates across borders.
Italy became the first EU member state to pass AI-specific legislation, effective October 10, 2025. The Italian law takes a position that the U.S. has avoided: AI-assisted works are copyrightable, provided they are the product of sufficient human intellectual work. The statute does not define “sufficient” in precise terms, which will inevitably require case-by-case adjudication, but it represents a fundamentally different philosophical starting point.
The EU AI Act entered into application on August 2, 2025, and includes transparency obligations for general-purpose AI model providers. They must identify and comply with rightsholders’ reservations of rights from text and data mining. The European Commission has launched a consultation on implementation details, and the Court of Justice of the European Union is expected to deliver its first ruling on AI and copyright in Like Company v. Google in late 2026 or early 2027.
The United Kingdom sits in an unusual position. The Data (Use and Access) Act 2025 received Royal Assent in June 2025 but did not change copyright law directly. Instead, it mandated that the government produce an economic impact assessment and a report on copyright policy options by the end of March 2026. The government’s preferred approach is a data mining copyright exception with a rightsholder opt-out, but an overwhelming 88% of consultation respondents opposed this, favoring mandatory licensing instead. Creative industries have been particularly vocal, arguing that an opt-out regime would effectively hand their work to AI companies for free.
| Jurisdiction | AI Authorship | Training Data | Key Legislation |
|---|---|---|---|
| United States | No (human authorship required) | Fair use if legally acquired | Case law (Thaler, Bartz) |
| European Union | Varies by member state | Transparency + opt-out rights | EU AI Act (Aug 2025) |
| Italy | Yes, with sufficient human input | Subject to EU framework | AI Act (Oct 2025) |
| United Kingdom | Undecided (review due Mar 2026) | Proposed exception with opt-out | Data Act 2025 + pending |
| China | Yes, if human involved in selection | Government-approved datasets | Interim Measures (2023) |
The Practical Problems Nobody Has Solved
Legal theory is one thing. Running a business is another. Companies that rely on AI-generated content face a set of practical questions that current law simply does not answer.
The first is originality detection. If a marketing team uses an AI tool to generate ad copy, and that copy closely resembles an existing tagline, who is liable? The AI company? The user? Both? Traditional copyright infringement requires proof of access and substantial similarity. But an AI model trained on billions of documents has, in a statistical sense, “accessed” everything. Whether that constitutes legal access under copyright law is a question no court has definitively resolved.
The second is ownership within organizations. If an employee uses a company-licensed AI tool to create a design, the work-for-hire doctrine likely assigns ownership to the employer. But if the AI-generated portions are not copyrightable, what exactly does the employer own? Only the human-selected arrangement? The overall compilation? The answer matters enormously for companies that need to license, sell, or enforce rights in their creative output.
The third is international enforcement. A work that is copyrightable under Italian law may not be copyrightable under American law. This creates incentives for jurisdiction shopping and makes cross-border licensing agreements nightmarishly complex. An AI-generated jingle used in a global advertising campaign might have different legal status in every country where the ad runs.
Content platforms are responding by hedging. Stock photography sites like Getty Images and Shutterstock have created separate AI-generated content tiers with different licensing terms and lower prices. Some publishers now require authors to disclose AI usage and accept reduced royalty rates for AI-assisted work. These are commercial workarounds, not legal solutions, and they vary wildly from platform to platform.
Where This Is Heading
Predicting the future of AI copyright law is difficult because the technology keeps outrunning the legal framework. But several trends are becoming clear.
First, the human authorship requirement in the United States is settled for now. After the Supreme Court’s refusal in Thaler, it would take either new legislation or a future Court willing to revisit the question to change this. Neither seems imminent.
Second, the “how you got the training data” question is emerging as the real battleground. The distinction Judge Alsup drew in Bartz, between legally acquired and pirated training materials, is likely to become the standard framework. This means AI companies will need to prove the provenance of their training data, and that proof will be expensive. Expect to see a growing market for “certified clean” training datasets with documented licensing chains.
Third, international fragmentation is accelerating. The EU, UK, and US are moving in different directions, and smaller jurisdictions are beginning to chart their own courses. For companies that operate globally, this means maintaining different IP strategies for different markets, a compliance burden that will particularly disadvantage smaller players.
The deeper question is philosophical. Copyright law was built on the assumption that creative works emerge from human minds. When a tool becomes sophisticated enough to produce work that is indistinguishable from human output, the premise itself is under strain. Whether the law adapts by broadening its definition of authorship or by creating an entirely new category of rights for AI-generated content, the U.S. Copyright Office’s ongoing AI initiative suggests the government is at least aware that the current framework cannot hold forever.
For now, though, the honest answer to “who owns what AI creates?” is: it depends on where you are, how the AI was used, how the training data was obtained, and which court you ask. That is not a satisfying answer. But it is the only accurate one.
Frequently Asked Questions
In the United States, only the portions that reflect your original human creative expression are copyrightable. If you wrote a detailed prompt and then substantially edited, arranged, or modified the AI output, those human-authored elements may qualify for protection. But purely AI-generated content with minimal human input cannot be registered. The Copyright Office evaluates each case individually, looking at how much creative control the human exercised over the final work.
The emerging legal consensus in the U.S. is that training on legally acquired copyrighted material can qualify as fair use, based on the June 2025 Bartz v. Anthropic ruling. However, using pirated or unlicensed copies does not receive fair use protection. In the EU, rightsholders can opt out of text and data mining under the AI Act’s transparency framework. The rules differ by jurisdiction, and several major cases are still pending that could further refine these boundaries.
Document everything. Keep records of your prompts, the edits you made, the creative decisions that shaped the final output, and the AI tools you used. Maximize your human contribution to any work you want to protect. For training data, ensure your AI providers can demonstrate lawful acquisition of their datasets. If you operate internationally, consult IP counsel in each jurisdiction where you plan to use or license the work, because the rules are diverging significantly across countries.