The two most common generative AI copyright mistakes are assuming that anything a model produces is automatically yours and protected, and assuming that using existing works as AI input or reproducing them in output is risk-free. These are separate problems with separate legal tests, and mixing them up is how many creators end up exposed on both fronts. U.S. copyright has long rested on human authorship, so the first question is whether a person made enough creative choices to claim the work. The U.S. Copyright Office’s January 29, 2025 report says AI-generated output may be protected when a human determines sufficient expressive elements, but simply supplying prompts is not enough by itself.
The second question concerns infringement, and that is where the stakes grow. A model trained on copyrighted material, or an output that closely resembles a protected work, can trigger claims from rightsholders. In the United States, statutory damages run from $750 to $30,000 per work and can rise to $150,000 for willful infringement. Courts in different countries are still working through these questions. Rulings from Munich and Delhi in July 2026 reached different outcomes on different facts, so a rule that holds in one place may not hold in another. Registration adds its own obligations: you must disclose more-than-de-minimis AI material and exclude it from your claim, and a Standard Application currently costs $65. Avoiding these mistakes starts with knowing which risk you are facing.
Key Takeaways
- You cannot assume that everything a generative AI model produces is yours to claim, because U.S. copyright requires a human to determine enough of the expressive elements, and supplying a prompt alone is not enough.
- When you register AI-assisted work with the U.S. Copyright Office, you must disclose any more-than-de-minimis AI-generated material, describe your human contributions, and exclude the generated material from your claim, and a Standard Application currently costs $65.
- Your infringement exposure can come from both the training data behind a model and the similarity between an output and a protected work, and U.S. statutory damages run from $750 to $30,000 per work, rising to $150,000 for willful infringement.
- If you publish or sell an infringing output, you can be liable even when a company built and trained the AI tool, so a vendor’s general assurance does not remove your own exposure.
- Your legal position depends on where you publish or sell, because courts in Munich and Delhi reached different outcomes in July 2026 on different facts, so you should check the rules in each country where your work will appear.
Why Does Authorship Matter for AI-Made Work?
Authorship matters for AI-made work because copyright is built on the idea that a work expresses a person’s choices. That means who made a work shapes both what you can own and what you might owe. It is a question about creative agency: who decided what the work would say, how it would look, and what would be left out?
In plain terms, authorship is the act of making a work through your own creative decisions. Originality is the quality that work must have to qualify for protection. It must come from independent effort and reflect at least a minimal creative choice, rather than copying something or mechanically following a template. Infringement is the unauthorized use of a protected work in a way that the law reserves to its owner, such as copying it, distributing it, or creating a substantially similar version of it.
Philosophers have argued about these ideas for centuries. John Locke grounded property in the labor a person mixes with what they produce. Hegel and later Romantic thinkers treated a work as an expression of its maker’s personality. Roland Barthes, in “The Death of the Author,” questioned whether a single originating mind needs to stand behind a text at all. Copyright law sits closer to the first two traditions. It protects expression that a human has shaped, not ideas or raw output that appears without a human guiding the result.
The two AI risks point in opposite directions because they answer different questions. The first concerns what you can claim. If your only contribution was a prompt, you may hold little or no protection in the output, which means others may be free to reuse it. The second concerns what you may be accused of. If a model was trained on works you did not license, or if your output reproduces a protected work too closely, a rightsholder may accuse you of infringement and seek damages. One risk asks whether you own something, and the other asks whether you owe someone else.
Keeping these apart changes how you work. A creator who asks only whether the output belongs to them may overlook a claim that a rightsholder could bring against them. A creator who focuses only on infringement may never record the creative decisions that give them a claim of their own. Each question calls for different evidence: a clear record of your choices and edits for the first, and a comparison of your output against existing works for the second.
Can You Claim Copyright in Something an AI Generated?

Usually not, at least not for the output as a whole. In the United States, you can claim copyright only in the parts of a work that reflect your own creative control. The Copyright Office has said that supplying a prompt, by itself, does not make you the author of what a system returns. The philosophical reason is that a prompt is closer to a commission than to hands-on making. When a patron asks a painter for a portrait of a particular sitter in a particular mood, the patron shapes the request, but the painter makes the many small decisions that turn that request into an image. A prompt sets an intention, while the model makes most of the expressive choices in between. Locke would ask who actually mixed their labor into the result, and a short instruction often leaves that answer unclear.
What can be protected is what you add yourself. Selecting and arranging generated material, writing or substantially rewriting text, editing an image in ways that create new expression, or combining several outputs into a new composition can all reflect human authorship. The test looks at whether a person determined enough of the expression, not at how long or detailed the prompt was. A lengthy, carefully specified prompt may still leave the expression with the system, while a brief prompt followed by your own substantial revisions may give you a claim in the revised work.
Some people have tried a different route by listing the AI system itself as the author. The dispute in Thaler v. Perlmutter tested whether a machine can be named as an author on a copyright application, and the Copyright Office’s position has been that a human author is required. For you, the practical lesson is that every author line on a registration must name a person, and your claim must rest on that person’s contributions.
Registration carries its own obligations. Under the Copyright Office’s March 2023 registration guidance, you must disclose any AI-generated content that is more than de minimis, describe the human-authored contributions you are claiming, and exclude the AI-generated material from the claim. If an omission is discovered later, it needs correcting. A failure to fix it can put the registration at risk of cancellation or cost you its benefits.
The cost of doing this properly is modest by comparison. Electronic registration for one work by one author who is also the claimant, where the work is not a work made for hire, costs $45. A Standard Application costs $65, and paper filing costs $125, according to the Copyright Office’s fee schedule. The fee is rarely the expensive part. The expensive part is discovering, after a dispute has started, that your application did not separate what you wrote from what the system produced.
Why Do Familiar-Looking Outputs Create Infringement Risk?
Familiar-looking outputs create infringement risk when a model copied protected works during training, or when its result reproduces recognizable parts of them. You can be exposed for publishing or selling that result even if a company built the tool. The first dispute concerns training. Rightsholders argue that copying their works to train a model is unlicensed reproduction, while developers argue that training is a lawful use. Courts have not settled this, and the answer depends on the case and the jurisdiction.
Germany offers one example. In July 2026, the Munich Regional Court I largely granted GEMA’s claims against Suno concerning six specified musical works, granting injunction, information, and damages. The court’s notice said song-lyric infringement was not part of that case. It also found that Suno’s training set included music copied from YouTube through stream-ripping that circumvented a technical protection measure, and that the works were reproducible in model versions v3.5 and v4 and recognizable in outputs.
India reached a different interim result. On July 24, 2026, the Delhi High Court dismissed ANI Media’s application for an interim injunction against OpenAI. At that stage, ANI had not shown substantial similarity in ChatGPT outputs or demonstrated memorization and regurgitation of its works. The court stressed that its observations only decided the interim application and would not determine the final outcome of the suit.
In the United States, developers often defend training by arguing fair use. Fair use is a defense, not a permission slip. A court weighs the facts of a particular use, including what was copied, how it was used, and what effect the result has on the market for the original. Someone must argue it, and the argument can fail.
The European Union handles text and data mining differently. Article 4 of Directive 2019/790 creates an exception for lawfully accessible works, but it does not apply when rightsholders have expressly reserved the relevant use in an appropriate manner. A publicly visible work can therefore still carry a reservation that blocks the exception.
The July 2026 Anthropic settlement shows how narrow such resolutions can be. On July 20, 2026, a federal court in the Northern District of California granted final approval to a $1.5 billion settlement covering only listed books downloaded from Library Genesis or Pirate Library Mirror, a list of 482,460 works. Reports put the allocation at about $3,000 per eligible book, but that is a settlement figure, not a damages rule for AI copyright cases. The release also does not cover claims based on model outputs or future conduct.
Output similarity is the risk you control most directly. A result can reproduce a distinctive melody, a memorable passage of prose, or a character’s signature details without you ever asking for that work by name. Rightsholders compare what the output looks or sounds like with the protected work, so your good intentions do not change the comparison.
In U.S. law, a copyright owner who wins can elect statutory damages instead of proving actual losses. Under 17 U.S.C. ยง 504(c), that election ranges from $750 to $30,000 per work, and a court may increase it to as much as $150,000 for willful infringement. The owner makes the election, and the count is per work, so a batch of many infringing images can multiply exposure quickly.
Downstream users carry this risk too. If you publish or sell an infringing output, the act of distribution is yours, so you can be liable even if a company built and trained the tool. A vendor’s general assurance does not remove your own exposure.
Generated code raises the same problem in a narrower form. If a code assistant returns a block that matches licensed or copyrighted source, and you ship it in a product, the copied code now sits in your product. Check generated snippets against existing code, and confirm which license terms apply before you reuse them.
Do Platform Terms Really Give You Ownership?

No. A platform’s promise to give you rights in its output does not by itself create copyright in that output. A terms-of-use agreement is a contract between you and the company, and it can only allocate what the company is able to grant, such as permission to use what the system returned. Copyright is different. It is a set of rights recognized against the world, and whether a work qualifies depends on the human authorship test that U.S. copyright applies, not on what a company says in its contract.
That gap is the source of much of the confusion. A sentence in the terms stating that “you own the output” is best read as a statement about what the company is willing to let you do with the result. It is not a finding that a court would recognize you as the author. Locke’s idea of property asks who mixed their labor into a work, and no contract can answer that question for you. A contract binds only the parties who signed it, while copyright can affect third parties who never agreed to anything, including someone who copies your output from a public website.
Terms of use often do several things at once. They may assign or license whatever rights the company holds in the output, restrict commercial use to certain subscription levels, prohibit using outputs to train competing systems, and limit the company’s own liability for errors or disputes. Read these as a list of promises the company makes and conditions it attaches to its service. They do not guarantee that your use is free of claims from people who own works the model learned from or resembles.
Risk-shifting clauses deserve particular attention. An indemnity clause requires you to cover the company’s losses, including legal costs, if an output you generated infringes someone else’s rights. A warranty clause may have you promise that the prompts, images, or documents you upload do not infringe anything. Together they can leave the company with a shield while you carry the bill, even though you had no say over the training data or how the model was built.
Terms also change. A company may update its agreement, and the version that governed the project you finished last spring may differ from the one posted today. Before you publish or sell anything, identify the terms in effect on the date you generated the work, and check whether the company publishes a revision history. Keep a dated copy or screenshot of each version you accepted, together with the plan you were on, because permissions often depend on account tier. If a dispute arises, that record shows what you were actually promised. It is far easier to assemble when you make it at the time rather than after a claim arrives.
How Do Court Rulings Differ Across Countries?
Court rulings differ across countries because each court applies its own copyright law to its own facts. A result in Munich or Delhi tells you little about how a similar dispute would end where you publish. Before relying on any ruling, ask where the work is sold, where the model was trained, and exactly what was copied.
The Munich decision turned heavily on how the training material was obtained. The court found that the training set included music taken from YouTube through stream-ripping, a method that gets around a technical protection measure. That detail matters for creators because a rightsholder’s case can depend on how the copies were made, not only on how closely an output resembles a protected work.
The Delhi ruling worked differently because it came at an early stage. Refusing an interim injunction means declining emergency relief before a full trial, and the judges were explicit that their observations decided only that application. A reader who treats an interim ruling as a final verdict has misread what the court said about its own decision.
The European Union takes a different approach to training data. Its text-and-data-mining exception covers lawfully accessible works, but a rightsholder can block it by expressly reserving the use in an appropriate manner. Text and data mining means automated analysis of large collections of material to find patterns, and the reservation is the rightsholder’s signal that such analysis is not permitted. The directive’s wording leaves room for argument about what counts as an appropriate reservation, so a creator working in Europe should not assume that silence on a website means permission.
The Anthropic settlement is often misread as a general ruling, and it is not one. A settlement is an agreement that a court approves to end a specific dispute, and its reach is set by the works it lists and the claims it releases. Here, that means a defined set of downloaded books. It cannot be read as a finding that all training on copyrighted books is lawful or unlawful.
Several other cases remain pending, including Getty Images v. Stability AI, in which the photo agency alleges that its images were used to train an image generator. Other suits against AI developers are still moving through courts in the United States, Europe, and elsewhere. None of these has produced an outcome that settles the question, so you should not assume how any of them will end.
For a working creator, the practical lesson is to treat each country where you publish or sell as a separate question. Note the jurisdiction, the date, and the terms of the tool you used, and check for local rulings before you rely on a precedent from somewhere else.
Is Adding Your Own Work Enough to Protect You?

Adding your own work can protect the parts you actually wrote or shaped, but it does not remove an infringing element that the output already contains. The human contribution and the infringement question are separate, and you need to answer both.
Contributions the law can recognize include selecting among several generated drafts, arranging passages or images into a sequence you designed, substantially rewriting sentences so the expression is your own, and adding original elements such as a new scene, a hand-drawn layer, or your own melody over a generated backing. The test is whether you determined enough of the expression in the final work. Changing one word or adjusting a color may add little. A collage built from ten generated images, each reworked and placed in an arrangement you chose, stands on firmer ground than a single image with a filter applied.
Your claim then covers only the human portions. In a registration, you describe those contributions and exclude the generated material, so the protected part must be specific enough that a reader can see where your authorship begins and ends. A vague description such as “AI-assisted artwork” does little for you. Naming the rewritten chapters and the arranged sequence gives a clear boundary.
Documentation is what makes that boundary provable. Keep the drafts you generated, the prompts you used with their dates, and an edit history that shows what you changed and why. Save versions as you work rather than trying to rebuild them from memory later, because a dispute may ask you to show which words or pixels were yours.
Infringement works differently. If the output reproduces a distinctive melody or a recognizable passage from a protected work, your later edits do not remove that copied element. Rightsholders compare the final work with the original, so a heavily revised version that still carries the copied material can remain infringing. The Munich court’s finding that the works were recognizable in outputs shows that the question is what the result still contains, not only who touched it last.
When you spot a matching element, your options are to remove it entirely, replace it with material you create from scratch, or drop the output. Adding a verse or a new background over a copied melody leaves the copy in place.
Building a Pre-Publication Check for AI-Assisted Work
Start with one piece you plan to publish this month, not your whole archive. Open a plain document and record the title, the date you generated the material, the tool and plan you used, and the version of the terms you accepted. Then list every human contribution you can name, such as the drafts you chose, the passages you rewrote, the layers you drew, or the melody you added. If that list is thin, treat the work as mostly machine output and plan accordingly before you register or sell it.
Next, check the generated parts against existing works. Run a reverse image search for pictures, search distinctive sentences in quotation marks for prose, and compare any melody or code block with material you can find. Note every match, even a partial one, and decide whether to cut it, replace it, or drop the piece.
Finally, write down each country where you will publish or sell, and look up whether its courts or regulators have addressed the issue. If the work will reach readers or buyers in the European Union, check whether any source you drew on carries a reservation of rights. Keep the completed sheet beside your saved drafts, and revisit it before each release, because the answers depend on the specific work, the tool, and the place it appears.
Frequently Asked Questions
1. Does a disclaimer saying a work was AI-assisted protect me from infringement claims?
No. A disclaimer does not change whether copied material appears in the final work, and rightsholders compare that work with the original they own. A recognizable melody, passage, or image taken from a protected source remains a problem even if you note that AI helped create it.
2. Is it safe to sell AI-generated images from a stock site I found online?
Not automatically. Read the license the site grants for that specific image, including whether it allows commercial sale in the way you plan. Then run a reverse image search to see whether the picture closely resembles an existing protected work, because the site’s permission does not settle whether someone else owns a copied element.
3. Can I use an AI tool to imitate a living artist’s style without risk?
Not without risk. Copyright protects particular expression rather than a general manner, so the real question is whether your output reproduces recognizable works by that artist, such as a distinctive composition or a copied passage. Prompts that aim to recreate a specific artist’s signature pieces invite the same resemblance comparison that applies to melodies and prose, and the outcome depends on the work and the jurisdiction.
4. Do the Anthropic settlement payments mean I can claim money for books I own?
No. The July 2026 settlement covers only listed books downloaded from Library Genesis or Pirate Library Mirror, and the list contains 482,460 works. Owning a book you bought does not put it on the list, and the reported figure of about $3,000 per eligible book is a settlement allocation rather than a general damages rule.
5. What counts as de minimis AI material when I register a work?
De minimis means a trivial amount, so the registration rule asks you to disclose AI-generated content only when it is more than that. Rather than relying on a fixed percentage, ask whether the system contributed a meaningful share of the expression in the work you are registering. If it did, describe the human contributions you are claiming and exclude the generated material.



