From code to canvas: the copyright fight over generative-AI creations.

Imagine a model drafts a novel — intricate plot, vivid characters, a bestseller. Who holds the copyright? The developer who built the model? The user who typed the prompt? Or no one, because the law sees no human at the keyboard of authorship? The intuition pump is fun, but the answer is now reasonably settled at one end and genuinely contested at the other. Untangling generative-AI intellectual property means separating two distinct fights: whether AI output can be copyrighted, and whether AI training infringed the works fed into it. They run on different statutes and they are moving at different speeds.

The output side: copyright needs a human author

U.S. copyright protects “original works of authorship fixed in any tangible medium of expression.”^[1] Courts and the Copyright Office have long read “authorship” to mean a human being — a thread that runs from a nineteenth-century photography case through monkey-selfie litigation to today. The generative-AI test case is Thaler v. Perlmutter. Computer scientist Stephen Thaler sought to register an image, “A Recent Entrance to Paradise,” that he said was generated autonomously by his “Creativity Machine,” which he listed as the sole author. The Office refused; in March 2025 the D.C. Circuit affirmed, holding that the Copyright Act requires a human author and that a machine cannot be one.^[2] The Supreme Court declined to hear the case in March 2026, leaving that holding in place.^[2]

It is important to read Thaler narrowly. It answers only whether a machine can be named as the author. It does not say that anything touched by AI is uncopyrightable. The harder, fact-specific question — how much human creative control is enough — is where most real disputes will live.

How much human input is “enough”?

The Copyright Office has been filling that gap. Its 2023 registration guidance told applicants that works may be registered when they contain sufficient human authorship, that purely machine-generated elements are not protected, and that applicants must disclose AI-generated material and disclaim it.^[3] In January 2025 the Office published Part 2 of its AI report, on copyrightability, concluding that existing law is flexible enough to handle generative AI and that no new statutory category is needed. Its line: prompts alone, however elaborate, generally do not make the user an author, because the user does not control how the system turns the prompt into expression. Protection attaches where a human contributes perceptible expressive choices — selecting, arranging, editing, or combining outputs with human-authored material — judged case by case.^[4]

For a working creator that is a practical instruction, not an abstraction. The copyrightable layer is the human one: the curation, the composition, the edits, the way generated pieces are assembled into something a person shaped. Keep records of that contribution. The generated raw material in the middle may sit in a gray zone or outside protection entirely.

The input side: the training-data suits

The second fight has nothing to do with who owns the output. It asks whether building the model — ingesting billions of copyrighted works as training data — was itself infringement, or whether it is fair use. Here the law is genuinely unsettled, and the cases are still in motion.

In Andersen v. Stability AI, a group of visual artists sued over images allegedly scraped into training sets. In August 2024 the court let the core direct and induced copyright-infringement claims proceed while dismissing certain DMCA claims; as of 2026 the case remains in active litigation and discovery.^[5] In The New York Times v. OpenAI, the publisher alleges its articles were used to train models that can reproduce its content; in April 2025 the court denied OpenAI's motion to dismiss the central copyright claim, allowing the case to move forward.^[6] Both remain pending — no merits verdict on whether training is fair use has issued in these cases.

Abroad, the picture diverges. In Getty Images v. Stability AI, the English High Court ruled in November 2025; after Getty dropped its primary copyright claims at trial, the court rejected the remaining secondary-infringement claim and made only narrow trademark findings.^[7] It is a UK judgment on UK law — useful context, not U.S. precedent.

What the Copyright Office says about training

In May 2025 the Office issued a pre-publication Part 3 of its AI report, on generative-AI training. It declined to create a new exception and said fair use must be assessed case by case under the existing four-factor test. Notably, it suggested that using vast troves of copyrighted works to produce content that competes in the same markets — particularly where access was unlawful — can fall outside fair use.^[8] The report is guidance, not binding law, and the courts will have the final word. But it signals where the analysis is heading.

What creators and counsel should know

A few durable takeaways survive the churn:

• Document the human layer. If you want registrable rights, preserve evidence of your creative choices — selection, arrangement, edits — and disclaim the purely AI-generated material when you register.^[3]
• Don't assume prompts confer authorship. Under current Office guidance, prompts alone generally aren't enough.^[4]
• Training liability is unresolved. Whether ingestion is fair use is being litigated, not settled; treat confident claims in either direction with caution.^[5][6]
• Provenance is a control, not just an aesthetic. Keep your generation logs, model versions, and source records — they are the evidence both registration and litigation will turn on.
• Jurisdiction matters. A UK or EU ruling does not bind a U.S. court, and the reverse is equally true.^[7]

The romantic question — can a machine be an author? — already has an answer in U.S. law: no. The interesting questions are quieter and more consequential. How much of you has to be in the work for the law to protect it, and was it lawful to feed the work to the machine in the first place? Those are being decided right now, one motion at a time.