Who Owns What AI Writes
AI and copyright — what's settled and what's still contested
① "Who owns what AI writes?" is really three separate questions — ⓐ does the AI output itself carry copyright, ⓑ what about the rights in the works used to train it, ⓒ what if the output you got copied someone else's work. The three answers differ in how settled they are.
② What's settled: both Korea and the US hold that "the author is human." Output an AI made on its own has no copyright — it belongs to no one — and only the part with human creative contribution is protected.
③ What's contested: whether mass-using text and images from the web to train an AI is copyright infringement is being litigated right now in 2026, and the answer may differ by country.
① Split it into three questions
"AI and copyright" is confusing because three different questions are packed into one phrase. Separate them first.
| Question | The issue | Status in 2026 |
|---|---|---|
| Copyright in the output | Does an AI-made result get rights | Mostly settled — only the human-contributed part is protected |
| Training data | Is mass-using works to train infringement | Contested — no final ruling |
| Infringing output | Did the result you got copy someone's work | Case by case — the user can be on the hook too |
This piece takes them in order. ② is the settled ground, ④ is the contested ground, ⑤ is the part a user has to watch personally.
② What's settled — the author is human
United States. Under copyright law the author must be a human. The US Copyright Office refused to register a picture that Stephen Thaler claimed "an AI created autonomously," the D.C. Circuit upheld that refusal (2025), and the Supreme Court declined to take the appeal on March 2, 2026, leaving it settled. The Copyright Office's 2025 report takes the same line — however carefully you write a prompt, that alone isn't enough, while the parts a human selected, arranged, or modified can be protected, case by case.
Korea. The Copyright Act defines a work as "a creative expression of human thoughts or emotions." An AI cannot be an author. The "Generative AI and Copyright" guides from the Ministry of Culture and the Korea Copyright Commission (first issued late 2023, two revised guides in June 2025) set out the same principle — registration is possible only where there is human creative contribution, and fully automatic output is not.
Both countries reach the same conclusion. Text an AI wrote in full is not "your work." If someone copies it wholesale, you can't stop them on copyright grounds.
Even with no copyright, platform terms of service, trademark, likeness/publicity rights, contractual disclosure duties, unfair-competition rules, and school or journal AI-use policies still apply separately. Don't jump straight to "not copyrightable, therefore free." Where the data you put into an AI goes is covered in AI and your privacy; the rights in what you get back sit at the same level.
③ Where does "human contribution" start counting
So how much do you have to do for it to be "your work"? The line is still drawn case by case, but the direction looks like this.
| What you did | Chance of protection |
|---|---|
| Only wrote an elaborate prompt | Usually not enough — a prompt is an "instruction," not an "expression" |
| Picked from many outputs (selection) | If the selection/arrangement is creative, possible at the compilation level |
| Substantial editing, additions, restructuring | The expression you edited can be protected |
| Human-written text + AI illustration | The human-written text is protected; the AI illustration is judged separately |
Why a prompt rarely counts as "expression" makes sense from how an LLM works — the same prompt yields a different result each time, and what fixes the final expression is the model's probabilistic sampling, not your sentence.
Use the AI as a tool, but make sure your judgment and edits materially shape the final version, and keep a record of what you changed and how. Both the US and Korea ask you, at registration, to identify which parts were AI-made and which were human-made. The workflow itself — "AI drafts, human finishes" — becomes a line of defense.
④ Contested — the training-data question
From here nothing is settled. The issue: is it copyright infringement for an AI company to collect news, books, images, and code at scale and train a model on them, or does it fall under fair use (US) or a country's exceptions?
The leading pending cases:
- The New York Times v. OpenAI and Microsoft — filed December 2023. It is in discovery in the Southern District of New York, with no trial date. It has been consolidated with similar suits from other news organizations.
- Getty Images v. Stability AI, class actions by visual artists, and other image-generation suits are also pending.
Three things are being fought over — ⓐ whether training is a "transformative use" rather than using the original as-is, ⓑ whether the model reproducing (regurgitating) substantial portions of an original is a separate infringement, and ⓒ whether the output substitutes for the market of the original.
As of 2026 there is no final ruling that settles this. And the answer may differ by country — the US has fair use, the EU has a text-and-data-mining exception with a rightsholder opt-out, Japan has a relatively flexible exception, and Korea is working on updating its rules. Treat any confident one-way claim with caution.
The outcome of these suits could translate into content-licensing spend, model-retraining cost, and settlement risk for Big Tech (Google, Microsoft, Meta) and AI startups. How to read regulatory and litigation risk into earnings is worth taking from the framework in the Big Tech earnings guide.
⑤ If the result you got copied someone else's work — user liability
Separate from the training-data question, sometimes the finished output looks too much like a specific work. Here liability for infringement can fall on the person who actually used the result. "An AI made it" is not a defense.
The risky types are well known.
- Images that closely imitate a famous character, a logo, or a specific artist's style
- Articles, lyrics, or code that comes out nearly identical to the original in substantial part
- Output made with prompts like "in the style of [artist X]"
Style itself is not copyrightable. But if the concrete expression is substantially similar to a specific work, it's a problem. Just as you verify "facts" an AI invents (the hallucination guide), it's safer to run anything you'll use commercially through a reverse-image search and a plagiarism check first.
⑥ A practical checklist (for users in Korea)
| Purpose | What to do |
|---|---|
| Blog / social content | Use the AI draft as a tool only. Polish the final expression yourself and check the platform's AI-disclosure rules. |
| Work documents / reports | Check company policy and the copyright-ownership clause in your contract first — especially for deliverables handed to a client. |
| Publishing / commercial work | Document the human contribution, and identify the AI-made parts separately at registration. |
| School assignments / papers | The institution's AI-use policy takes priority, separate from copyright. Check what's allowed first. |
| Images / design | Run a similarity check before commercial use. Check the generation tool's terms for commercial rights and ownership. |
The practice of writing with AI itself is covered in writing with AI. This piece adds the rights picture for what you produce that way.
One thing is settled — the author is human, and output an AI made in full has no copyright. For it to count as yours, you have to leave a material contribution through selection, editing, or combination. Whether training data is infringement is still being litigated, so filter out the confident claims, and check similarity and licensing before any commercial use.
※ Written as of September 2026, describing general principles and current status. This piece is not legal advice. Copyright outcomes depend on the country, the case law, and the specific facts, and the issues described here as "contested" can change with later rulings. For an actual dispute or a registration/contract question, consult a lawyer or a body such as the Korea Copyright Commission.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.