Who owns AI's words and pictures?

You type a few words into an AI tool. Out comes a short story, or a picture, or a song. Two questions arrive with it, and the law is still arguing about both.

First: the model learned to do this by studying millions of things other people wrote and drew. Were those people owed anything? Second: the new thing on your screen, the one you shaped with your words, is it yours to own?

This is the copyright lesson. Lesson 6 put ownership on the risk map as one square big enough to earn its own lesson. Here it is. We will map the fight honestly, as a debate to understand, not a verdict to hand down, so you can follow the argument in the news without being pushed around by either side.

Why we protect ideas at all

Start with a strange fact. You can own a chair. Can you own a song? A song is nothing you can hold, yet long ago the law decided some creations of the mind can be owned anyway. We call them intellectual property: writing, music, art, inventions, ideas that the law treats a little like property. Copyright is the branch that covers creative expression, the actual words, images, and sounds a person makes.

Why grant this kind of ownership at all? The Harvard Kennedy School course this track adapts, created by Sharad Goel, Dan Levy, and Teddy Svoronos, gives two answers.

The main one, and the one that drives United States law, is a bargain. Society hands a creator a limited, temporary head start over what they made. For a while, others need permission to copy it. That lets the creator earn from the work, which gives people a reason to keep creating. The protection does not last forever. When it ends, the work enters the public domain, which just means anyone can use it freely. So copyright is a deal struck for a purpose: more creation, for everyone, over the long run.

The second answer is older and simpler. What I make is mine, and taking it without asking feels like theft. The course calls this the natural-rights view. It carries less weight in United States law, but it shapes how people feel about the fight.

Did the model learn from work it never paid for?

Here is the first hard question, and the one with the most money riding on it. To build a modern AI model, its makers fed it an enormous pile of human writing and images. A lot of that material was copyrighted. Nobody asked the creators. Was that infringement?

Two plain definitions first. Infringement means using someone’s protected work, without permission, in a way the law forbids. Fair use is the escape hatch: a doctrine in United States law that allows some use of copyrighted work without permission, so copyright does not choke off things like teaching, news, research, and commentary.

How that training works is a different track; How models are pretrained covers it. Here we need only that it happened.

The course’s 2024 centerpiece was a lawsuit. The New York Times sued OpenAI and Microsoft, saying they trained on millions of Times articles without permission. The companies agreed they had used the articles, and argued it was fair use. As of July 2026, that case is still going. A judge let much of it move forward in 2025, but there has been no ruling on the core question and no trial date. It is unresolved.

To weigh a fair-use claim, United States courts look at four things.

The purpose of the use. Is the new use different in kind from the original, or does it just repackage it? A use that adds something genuinely new is called transformative, and that weighs toward fair use.
The nature of the work. Plain facts get thinner protection; creative work gets more.
The amount used. Taking a little weighs toward fair use; taking the whole thing weighs against it.
The effect on the market. This one often decides. Does the new use compete with the original and eat into its income?

Now the fresh part, because since 2024 real courts have actually ruled. Read these for exactly what they are: narrow, fact-specific decisions, each from a single trial court, none yet ruled on by a higher court.

In February 2025, in Thomson Reuters versus Ross Intelligence, a Delaware court decided that a startup’s copying of a legal-research company’s material, used to build a rival tool, was not fair use. This is a decided ruling, but a limited one. The tool was not a generative AI. The decision turned on its own facts. And it is under appeal; a federal appeals court heard arguments in 2026, with a decision still to come.

In June 2025, two California courts ruled within weeks of each other, and they did not fully agree. In Bartz versus Anthropic, the judge decided that training an AI on books the company had lawfully bought was fair use, and called that use highly transformative. But the same judge decided that downloading and keeping pirated copies of books was not fair use. That part was infringement. In Kadrey versus Meta, a different judge ruled for the AI company on the training question, then went out of his way to warn that he was ruling narrowly. The people suing, he said, had simply argued their case poorly, and in most future cases he expected this kind of training to be found infringing, not fair.

Notice what just happened. One court said not fair use. Two said a particular training use was fair use, and both hedged hard. The rulings conflict, they are narrow, they come from trial courts with one already under appeal, and no higher court has settled the matter. So the honest summary is the one the course gave in 2024, still true in 2026: whether training AI on copyrighted work counts as fair use is an open question. Do not let anyone tell you it has been answered, in either direction.

One more, kept deliberately separate. After its ruling, Anthropic reached a deal to settle the piracy part of the Bartz case, agreeing to pay at least 1.5 billion dollars and to destroy its pirated copies. A class settlement like this needs a judge’s approval, and this one went before a court for that review in 2026. A settlement is a deal to end a case. It is not a court finding, and settling is not an admission of anything. We report it as a fact, and draw no lesson from it about who was right.

Other suits are working through the same California court, brought by visual artists and by a large image library, all unresolved as of July 2026. Separately, a related case in the United Kingdom reached a decision in late 2025 that mostly went against the image library. That ruling sits under a different country’s law, does not bind United States courts, and is itself being appealed.

For a creator worried their work helped train a model: the stakes are real, and you are not owed a settled answer yet.

Can what AI makes be owned, and by whom?

Now flip to your side of the screen. You prompted the tool. You have a picture. Can you copyright it?

The bedrock rule is human authorship: United States copyright protects work made by a human being. The reason runs straight back to the bargain. Copyright exists to give people a reason to create. A machine does not need that reason, so a work with no human author has nothing the deal is meant to protect.

This one is settled at the top, for one clear situation. A computer scientist tried to register an image he said his program had made entirely on its own, listing the machine as the author. The Copyright Office refused. He sued. In 2025 a federal appeals court agreed with the Office: under the Copyright Act, an author has to be a human. In 2026 the Supreme Court declined to take the case, which left that ruling in place. Notice the limit, though. That case was about a work claimed to be made with no human involvement at all. It did not decide the harder, far more common situation.

Which is this. You did involve yourself. You wrote the prompts, picked among the results, maybe edited the image. Is that enough human authorship to earn copyright? That line is exactly where the fight now sits. One well-known case involves an artist who made a prize-winning image with an AI tool after hundreds of prompts and revisions, then was refused registration. He is challenging that refusal in court, and as of July 2026 there is no final ruling. Unresolved.

While the courts work, the Copyright Office has told the public how it will handle registrations today. This is the Office’s own practice, not a final word from the courts, but it is the clearest guidance we have. Its position, set out in a 2025 report: typing prompts alone does not make you the author, because you are not steering the actual expression closely enough. The same tool can turn one prompt into many different pictures. But your own creative choices can be protected. If you select, arrange, and edit AI output in a creative way, or feed in your own writing that shows up in the result, that human part can be copyrighted, even though the raw AI-made part cannot. The Office registered one comic book on exactly this basis: the arrangement of the panels and the human-written words were protected, the AI-made images on their own were not.

So the practical answer for a user is a split. The bare output of a prompt is probably not yours to own; the creative work you did with it quite possibly is. The more of you that ends up in the final piece, the more there is to protect.

And when a piece is copyrightable, who holds the right? The course names three candidates: the person, the company that made the AI, or the AI itself. The AI is out, for the reason above. Some companies, through their terms of service, hand any rights in the output to the user, but terms differ from tool to tool and change over time, so this is one to check, not assume.

When the output copies too closely

One last question. Can an AI’s output break copyright by copying someone else’s work?

Back in 2023, a track called Heart on My Sleeve went viral. It was AI-made, but it sounded like two famous musicians who had nothing to do with it. Streaming services pulled it down. So, did it break copyright?

United States law asks two questions. Did the tool have access to the original work? For most large models the answer is almost certainly yes; their training swallowed huge amounts of published material. And is the new work substantially similar to a specific original? That second question is the crux, and it turns on a distinction worth keeping.

Copying a specific work is barred. Imitating a general style is not. You are allowed to write a song in the style of an artist you admire; human songwriters do this all the time. What you cannot do is reproduce a particular song. By that standard, a fake in someone’s style, like the viral track, probably does not break copyright, even though it borrows the artist’s sound. A new work built directly on top of a specific protected work is called a derivative work, and making one needs the original owner’s permission. A work that is merely in a similar style is not a derivative work.

There may be other laws in play. Many places recognize a right of publicity, which protects a person’s name, voice, and likeness from being used without consent. A voice clone could run into that even when it does not break copyright. But that is a different rule, aimed at a different harm.

Here the course raises the sharpest question and leaves it open on purpose. These style rules were written for humans, who find it slow and hard to copy another artist well. An AI can imitate almost anyone, almost perfectly, at scale. Should style still be the line when the copying is this easy and this good? Thoughtful people disagree, and the law has not answered. As of July 2026, this one is unresolved too.

Why this matters when you use AI

Follow the bargain. When you meet a copyright headline, ask what deal each side wants, not just who is angry.
Tell a ruling from a rule. One court deciding one case on its own facts is not settled law. Anyone who tells you it is decided is selling something.
Know your split. The raw output of a prompt is probably nobody’s to own. The creative work you do with it can be yours.

Common pitfalls

Treating one court’s decision as the law of the land. A single trial court, ruling one case on its own facts with an appeal pending, is a data point, not a nationwide verdict.

Reading a settlement as a confession. A company can settle to make a risk go away and admit nothing.

Assuming the prompt makes it yours. Typing a request is not authoring a work. What you can own is the creative human part you added, not the bare output the tool handed back.

Mistaking style for a copy. Sounding like an artist is generally allowed. Reproducing their specific work is not.

What you should remember

Copyright is a bargain: a limited, temporary head start meant to encourage more creation, after which the work enters the public domain.
The training question, whether it is fair use to train AI on copyrighted work, is genuinely open. Courts have ruled both ways in narrow, fact-specific, appealable cases, and no higher court has settled it.
Fair use turns on four things: the purpose of the use, the nature of the work, the amount used, and the effect on the market.
Purely AI-made work is not copyrightable, but your own creative contribution can be. The more human input, the more protection.
Copying a specific work can infringe; imitating a general style generally does not, though other laws like the right of publicity may still apply.

What’s next

You now know the state of the fight over who owns what AI makes and learns from. The next lesson turns from who owns the work to who does the work. If a model can write, draw, summarize, and analyze, what happens to the people paid to do those things? For many readers that is the most personal question in this whole track. We will meet it with the same care we brought here.

If you remember one thing

Copyright is a bargain, not a wall: a limited deal meant to keep both people and machines making good things.
The big question, whether training on copyrighted work is fair use, is honestly unsettled, and no ruling yet has answered it for good.
What a prompt hands back is probably no one’s; what you shape with your own hands can be yours.