How do copyright rules work for AI-generated content?

Short answer: in the United States, a work with no human author gets no copyright. Typing a prompt — even a long, clever, iterated one — does not make you the author of what the AI produces. Using AI as a tool inside a human creative process is fine. Using it as a stand-in for your creativity is not protected.

That is the settled part. Around it sits a live, expensive fight about the other direction of the same question: whether AI companies were allowed to train on your work in the first place. The two questions sound similar. The law treats them as completely different problems.

The bedrock rule: copyright needs a human

American copyright protects "original works of authorship." For over a century, "authorship" has meant a human being. The doctrine that sank the famous "monkey selfie" case — a macaque took the photo, so nobody owned the copyright — applies just as neatly to a diffusion model. The machine is a tool, not an author.

The US Copyright Office has now said this repeatedly and in detail, most fully in Part 2 of its Copyright and Artificial Intelligence report, published January 29, 2025. After reviewing more than 10,000 public comments, the Office concluded that existing law already answers the question and no new legislation is needed: works generated wholly by AI sit outside copyright, full stop.

This is not a quirk of American law. The principle runs deep enough that when the same question was asked in patent law — can an AI be listed as an inventor? — the US and UK systems gave the same answer: rights attach to people. Two different areas of law, asked essentially the same question, returned the same answer.

The case that tested the extreme

The cleanest test came from Stephen Thaler, a computer scientist who tried to register an image called "A Recent Entrance to Paradise" with his AI system — the "Creativity Machine" — named as the sole author and himself listed merely as the owner. The Copyright Office refused. The courts backed the refusal. The D.C. Circuit affirmed in 2025 that a copyrightable work must be authored by a human being, and in 2026 the Supreme Court declined to hear the appeal.

That sequence matters. The D.C. Circuit was careful to say it was applying the statute as written, and that any expansion of authorship to include machines would have to come from Congress. The Supreme Court's refusal to take the case left the human-authorship rule firmly in place. There is now no serious ambiguity in the US about the extreme case: zero human authorship, zero copyright.

Thaler's case was deliberately extreme — he wanted the AI named as author, which almost nobody actually wants. The interesting questions live in the middle, where humans and machines collaborate. That is where the Copyright Office drew its real line.

Where the line actually sits: tool vs. stand-in

The 2025 report's central distinction is between using AI as a tool and using AI as a stand-in for human creativity. The distinction depends on how the system is used, not on the technology itself — and it is analyzed case by case.

AI as a tool does not destroy copyrightability. If you write an essay and use AI to polish sentences, the essay is yours. If you shoot photographs and use AI to remove noise, the photographs are yours. If AI-generated material appears inside a larger human-authored work — a human-written book with AI illustrations, a human-edited video with AI b-roll — the human-authored parts remain protected.

What does not work is the prompt. The Office was explicit: prompts, regardless of complexity or specificity, do not amount to authorship. Even iterative prompt refinement — generating, judging, regenerating — was deemed insufficient, because the AI system independently determines the final expressive elements. Effort is not authorship; copyright law rejected the "sweat of the brow" doctrine decades ago, and it is not coming back for prompt engineering.

The early case that sketched this middle ground was Zarya of the Dawn, graphic novelist Kris Kashtanova's 2023 registration fight. The Office's decision was surgical: the human-written text was copyrightable, the human's selection and arrangement of images counted as authorship — but the AI-generated images themselves were not. You can own the collage without owning the pictures. That remains the best mental model for hybrid work: copyright follows the human decisions, not the machine's output.

One procedural note with real teeth: since 2023, the Copyright Office requires applicants to disclose AI-generated content in works submitted for registration, and to claim only the human-authored material. Registering an AI-heavy work without that disclosure is a good way to own nothing and learn it at the worst possible moment.

The other fight: your work training their models

Everything above is about outputs — who owns what the AI makes. The other direction is inputs: was it legal for AI companies to train on your work at all? That fight is bigger, messier, and still unresolved.

The Copyright Office addressed it in Part 3 of its report, released as a pre-publication version on May 9, 2025. Its conclusion was deliberately not categorical: fair use for training is fact-specific. Training a model will often be transformative — but, in the report's key sentence, "making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, goes beyond established fair use boundaries."

Translation: training a research model on lawfully acquired text is one thing; vacuuming pirated books to build a commercial product that competes with the authors is another. The report declined to recommend new legislation, preferring to let the licensing market develop — with a warning that targeted intervention could follow if the market fails specific kinds of creators.

The politics around the report were extraordinary. The Librarian of Congress was dismissed the day before its release; the Register of Copyrights, Shira Perlmutter, was fired the day after. In 2026, the administration filed briefs backing a broad fair-use reading for AI training in the OpenAI litigation — a sharp turn from the report's fact-specific approach. Meanwhile the lawsuits grind on: the New York Times' case against Microsoft and OpenAI, consolidated into multidistrict litigation, and a publisher class action against Meta alleging training data was deliberately sourced from pirate "shadow libraries." Discovery in these cases has surfaced internal messages — including a Microsoft researcher's description of the practice as "an astonishing theft of unprecedented proportions" — that will be quoted in courtrooms for years.

For creators, the practical upshot of the training fight is mostly defensive: your published work has very likely been ingested into training datasets already, and no individual opt-out mechanism has proven meaningful at scale. Collective licensing — the report's suggested fallback — remains more discussed than built.

What this means for you, practically

Strip away the case law and the politics, and the working rules for a creator using AI in 2026 are straightforward:

• Raw AI output is not yours in any legally meaningful sense. Do not build the crown jewel of your business on an asset you cannot own — a brand mascot, a flagship course, a book — generated start-to-finish by a model.
• Your human contribution is ownable. Your text, your edits, your arrangement, your selection, your modifications: document them. If a dispute ever comes, "I made these specific creative decisions" beats "I directed the AI" every time.
• Disclose AI use in copyright registrations. The Office requires it, and undisclosed AI material can poison an otherwise valid registration.
• Read your AI tool's terms, but don't confuse them with copyright. A platform granting you "commercial use" of outputs is a contract permission, not a copyright. It doesn't stop someone else from generating something similar, and it doesn't give you the exclusive rights that copyright would.
• For client work, put AI use in writing. If a client assumes they're buying full copyright in a deliverable that's substantially AI-generated, that misunderstanding becomes your problem the day they try to enforce it.

The honest uncertainties

Plenty remains genuinely unresolved. How much human modification is "enough" to claim authorship over AI-assisted work? Nobody knows — it's case-by-case by design. Will Congress legislate, either to expand authorship or to create a licensing regime for training? The Copyright Office said no new law is needed for outputs; the training question is politically live. How will courts rule in the mega-lawsuits? Years away, and any single ruling will be appealed.

But notice what is not uncertain: the core rule. Human authorship is required. Prompts aren't authorship. Tools are fine; stand-ins aren't protected. That much is settled enough to build on.

There is a quiet irony in all of this that creators should sit with. The same technology that lets anyone generate a passable image in seconds is governed by a legal principle that says: the valuable part was never the image. It was the human decisions inside it — what to keep, what to cut, what to mean. Copyright, asked to reckon with machines that can produce, ended up drawing a circle around the one thing machines can't do: decide. That was always the job. The law just finally said so out loud.