Using AI images legally: rights, credit and the safe path

Two questions get mixed together here, and separating them removes most of the confusion. The first is whether you are allowed to use the image, which is answered by the terms of the tool that made it. The second is whether you own it, meaning whether you can stop anyone else from using it, which is answered by copyright law and in several countries the answer is no. You can be perfectly entitled to put a generated image on a product and still have no way to object when a competitor lifts it. Add to that the things a model will happily draw and you may not publish, such as a trademark or a recognizable face, and you have the whole picture.

The tool's terms decide what you may do

Read the terms before you read anything about copyright, because that is the document binding you directly. Generators vary more than people expect, and the differences sit in predictable places.

• Assignment of output. Most consumer services say they assign to you whatever rights in the output they may hold. That is a careful sentence. It transfers whatever exists, which in some countries is nothing.
• Commercial use tied to the plan. Free tiers frequently permit personal use only, or grant everyone a broad license to reuse images made on that tier. Paid tiers usually unlock commercial use and privacy.
• Use restrictions. Common ones include no training a competing model on the outputs, no impersonating real people, no medical or legal advice imagery, and content rules you agreed to when you signed up.
• What happens to your uploads. If you supply a reference photo, check what license you grant over it and whether it can be used to improve the service.
• Indemnity. Some business tiers promise to defend you if an output leads to a copyright claim, subject to conditions such as not disabling their filters. Consumer tiers rarely offer this.

Terms change, and they bind the version you accepted only until you accept the next one. Keep a dated copy of the terms for any image you use commercially. If you are buying access for a team, the clauses that matter and the ones people forget to negotiate are covered in what to check in an AI vendor contract.

Why the US will not register an AI image

United States copyright protects works of human authorship. The Copyright Office has taken the consistent position that material produced by a generative system in response to a prompt is not human authored, on the reasoning that a prompt influences the result without controlling how it is expressed. A registration for a comic book made with a generator was accordingly limited: the written text and the selection and arrangement of the images were protected, the individual images were not.

What follows from that in practice matters more than the theory.

• You cannot register the raw output, and registration is a precondition for suing over infringement of a US work.
• Your human contribution can be protected. Meaningful editing, compositing, retouching and the arrangement of images in a larger work are authorship, and you register those while disclaiming the generated material.
• A client who needs to own artwork outright, for a logo or a character they will build a brand on, is not getting that from raw generated output. Treat generation as a stage in the process rather than the delivery.

The line sits at human creative control over expression, which is why a heavily reworked composite stands on firmer ground than a first generation straight out of the tool. If you want the mechanics behind why a prompt steers rather than specifies, see how a generator turns noise into a picture.

The UK and EU positions are different

The UK provision for computer generated works predates this technology and was written with things like procedurally created graphics in mind. It has barely been tested on modern generators, the government has consulted on whether to keep it, and nobody should build a business on it applying the way they hope.

The European position flows from the requirement that a work be the author's own intellectual creation, expressing free and creative choices. Typing a prompt and picking a favorite from four results is a thin basis for that claim. Directing, masking, compositing and reworking is a much better one. The EU also layers transparency duties on top, which are about labeling rather than ownership, and are discussed below.

Training lawsuits are a separate fight

The litigation you read about is largely between rights holders and the companies that built the models, over whether collecting and training on copyrighted images without permission was lawful. That dispute is unresolved in most countries and is being argued differently in each. It is not, for the most part, your dispute. Losing it would hurt the model makers rather than everyone who ever generated a picture. Where it came from, and why the arguments are so hard to settle, is covered in where AI training data comes from, with the copyright arguments themselves in AI and copyright.

Your exposure is narrower and more specific: publishing an output that is substantially similar to an existing protected work. That happens most often when you ask for it, by naming a living artist, a studio style, a franchise or a well known image, and it happens occasionally by accident when a training set contained the same picture thousands of times. Prompting for a named style is where people most often walk into trouble, and it is avoidable. Describe the qualities you want, light, palette, medium, era, rather than borrowing a name, which is the approach set out in writing image prompts that describe what you actually want.

Trademarks, faces and the rules that still bite

These catch far more people than copyright does, because a generator will produce them on request and nothing warns you.

• Trademarks and trade dress. A logo on a cup, a distinctive bottle shape, a store front. Trademark law is about consumer confusion, so the risk depends on how you use it. Editorial and clearly parodic use is treated differently from putting it on your packaging.
• Characters. Famous characters are protected independently of any single picture of them. Generating your own drawing of one does not make it yours.
• Real people. In the US a right of publicity protects name and likeness commercially, and it varies state by state, including how long it lasts after death. In the UK there is no standalone image right, so claims run through passing off and advertising rules. In the EU many countries have explicit image rights, and a realistic depiction of an identifiable person may also raise data protection questions.
• Advertising standards. Regulators generally require that advertising not mislead about a product. A generated image showing a result the product cannot deliver is a problem no disclaimer solves.
• Sensitive contexts. News, courts, medicine and finance have their own norms about illustrative imagery, and passing off a synthetic image as documentary costs more in credibility than in law.

Who you have to tell

Disclosure obligations come from four separate directions and they do not overlap neatly. Contracts come first: many client agreements, stock libraries, competitions and publishers now require you to declare AI generated material or forbid it outright, and breaching that is a contract problem regardless of copyright. Platform rules come second, with several large services asking uploaders to label synthetic media and applying their own labels when they detect it.

Third, regulation. EU transparency rules require providers of generative systems to mark synthetic output in a machine readable way and require realistic fakes of people or events to be disclosed. Fourth, ordinary honesty about whether an image depicts something real, which is what audiences actually care about. The technical layer under most of this is the signed provenance record described in how Content Credentials work, and keeping it intact through your editing process is easier than reconstructing it later.

A checklist before you publish

1. Confirm your plan permits commercial use, and save a dated copy of the terms.
2. Reverse image search the final picture. If it is close to an existing work, regenerate rather than argue about it.
3. Remove stray trademarks, logos and half formed brand names from backgrounds.
4. Check no one in the image is recognizable as a real person, and never use a real face without written permission.
5. Check the caption. The claim you make about what the image shows is where deception actually happens.
6. Keep the prompt, the seed, the tool and the date. If ownership or originality is ever questioned, that record is the only evidence you will have.
7. Label it where a contract, a platform or plain honesty calls for it.
8. For a logo, a mascot or anything a brand will be built on, commission a human or have a human rework it substantially, so that there is something to own.