How Does Copyright Law Relate to AI Generated Images?

The Human Authorship Problem: Your AI Image May Be Legally Ownerless

If you're using AI-generated images in your marketing, branding, or products, there's a possibility you need to confront: under current U.S. law, you may own nothing at all. Courts have made clear that "human authorship is a bedrock requirement of copyright," meaning images produced solely by AI prompts generally fall outside copyright protection entirely.

The pivotal case is Thaler v. Perlmutter. A D.C. district court affirmed the Copyright Office's refusal to register "A Recent Entrance to Paradise," an image authored by the "Creativity Machine" AI, holding that only human beings qualify as authors under U.S. copyright law. The D.C. Circuit later affirmed that decision, denied rehearing, and the U.S. government filed a brief opposing Supreme Court review — so this is settled law for now, according to the Copyright Office's AI resources.

Prompts alone won't save you. The Copyright Office's position is that written prompts do not provide sufficient human control to make users authors of an output, because users do not oversee, direct, or understand what the model contributes. There's also a practical "black box" problem: identical prompts can produce different outputs, demonstrating that AI cannot yet translate an idea into fixed, tangible expression.

The business consequence is stark. A competitor could take your prompt-only hero image, product visual, or ad creative and reuse it freely — and you would have no legal recourse, because there is no copyright to enforce.

What actually creates protectable ownership? Human creative contribution. The Copyright Office analyzes this case by case, looking at factors like:

• Selecting and arranging AI-generated material in a sufficiently creative way
• Modifying AI outputs enough to meet the originality standard
• Treating AI like a traditional tool, where what matters is "the extent to which the human had creative control over the work's expression"

The Zarya of the Dawn decision illustrates the line: the Copyright Office granted copyright to the comic book as a whole but refused protection for its individual Midjourney-generated images, ruling they were not sufficiently produced by the artist. IP lawyer Scott Hervey summarizes the rule plainly: "Copyright will only protect the human-authored aspects of the work."

One more trap to avoid: if you do register a copyright in an AI-assisted work, you have an affirmative duty to disclose the AI-generated material. Applications must name only human authors with a general statement noting AI content — and failure to disclose risks the registration itself.

The practical takeaway for any business leaning on AI visuals: assume prompt-only images are legally ownerless, and build in meaningful human editing, arrangement, and documentation of your creative choices if protection matters to you. As the legal landscape continues shifting — the Copyright Office issued its Part 2 report on generative AI output copyrightability in January 2025 — documenting your human contributions is the one step that never hurts. Agent skills can help automate that documentation workflow, and you'll be able to browse and test free options on SkillzBench.

When AI Images CAN Be Copyrighted: The Human Contribution Test

Here's the good news: using AI in your creative process doesn't automatically disqualify your work from copyright protection. The U.S. Copyright Office draws a sharp line between works generated by AI and works created with AI — and the second category can absolutely qualify.

The landmark example is Zarya of the Dawn, a comic book whose artist used Midjourney for its images. In its February 2023 ruling, the Office granted copyright to the comic as a whole — the story, text, and arrangement of images — but refused protection for the individual Midjourney-generated images themselves, as reported by The Art Newspaper. The human-authored elements were protectable; the raw AI outputs were not.

The Office's March 2023 guidance, "Copyright Registration Guidance for Works Containing AI-Generated Materials," clarified the test: a work containing AI-generated material can still qualify if a human selects or arranges the AI material in a sufficiently creative way, or modifies it enough to meet the standard of protection. The guiding principle is creative control — the extent to which a human directed the work's expression, treating AI much like Photoshop or a guitar pedal.

In practice, that means creators should focus on three protectable contributions:

• Selecting AI outputs and arranging them creatively into a larger composition
• Significantly editing or modifying AI-generated images
• Contributing original human-authored elements — text, layout, narrative — around the AI material

One critical responsibility comes with registration: applicants must disclose AI-generated material beyond minimal use, and applications must name only human authors. According to RAND's policy analysis, the Office relies on these disclosures to distinguish AI-assisted works from purely AI-generated ones — and failure to disclose can put the registration itself at risk.

Is this achievable in practice? Yes. As of February 2024, the Copyright Office had issued registrations to well over 100 AI-assisted works, per the same RAND report. The path exists; it just requires genuine human contribution and honest disclosure.

For teams building AI-driven creative workflows, documenting your human edits and disclosure steps matters as much as the creative work itself. Agent skills can automate parts of that pipeline — image processing, edit tracking, disclosure checklists — and SkillzBench (currently in pre-launch, with a live waitlist) is building a free, security-scanned directory where you can discover and benchmark exactly those kinds of skills before deploying them.

The Other Side of the Risk: Training Data Lawsuits and Output Similarity

So far, the question has been "can I protect my AI-generated image?" But there's a darker question every business should ask first: "could this image get me sued?" The answer, increasingly, is maybe.

The courtroom battles started with the training data itself. In Andersen v. Stability AI, artists allege that Stability AI, Midjourney, and DeviantArt scraped billions of images to train their tools without permission, and the court found the allegations sufficient to plead direct infringement (per legal analysis of the case). Meanwhile, Getty Images opened its landmark trial against Stability AI before the British High Court in June 2025 — widely described as the first major copyright trial of the generative AI industry.

Getty, which represents nearly 600,000 artists globally, alleges Stability AI "unlawfully copied and processed millions of images protected by copyright and the associated metadata" without a license. Getty's trial lawyer put it bluntly: Stability "didn't care if images were protected by copyright, had watermarks, were not safe for work, or were pornographic — it just wanted to get its model to market as soon as possible."

The unresolved fair use question sits at the center of all this. Whether training models on copyrighted works is fair use remains undecided, and generative AI complicates the traditional "non-expressive use" defense: outputs can reproduce particular works with high similarity and may directly compete with the works they were trained on, as RAND policy research notes. More than two dozen lawsuits are now in flight.

Then there's the output problem — the one that lands on your desk. In experiments published in IEEE Spectrum, Gary Marcus and Reid Southen found Midjourney could generate images that "appear to display copyrighted material," framing the results as a copyright minefield. Stable Diffusion has even produced images with Getty watermarks still visible. If your AI tool can cough up someone else's trademark or a recognizable film character, publishing that output is your problem.

For commercial use, screen every AI-generated image before it ships:

• Check for recognizable copyrighted characters, logos, or watermarks — even faint ones
• Run reverse image searches to catch near-duplicates of existing works
• Document your prompts and human edits, in case provenance is questioned later
• Review your AI tool's terms of service on output ownership and indemnification

These checks are exactly the kind of repetitive workflow that agent skills can automate — SkillzBench is building a free, verified directory of them, with a waitlist open now. Until courts settle fair use, assume your image is only as safe as your screening process makes it.

A Global Patchwork: Why the Rules Change at the Border

An image that earns copyright protection in London might be unprotectable in Washington, D.C., and freely monetizable by anyone else in Beijing. For businesses publishing or selling visuals internationally, that patchwork is not an academic curiosity — it directly affects who owns what.

The United States sits at one end of the spectrum. Under U.S. Copyright Office guidance, human authorship is a bedrock requirement, and courts affirmed in Thaler v. Perlmutter that purely AI-generated images cannot be registered. Prompts alone do not make a user an author.

The United Kingdom takes the opposite approach. British law provides statutory protection for computer-generated works, meaning an image produced without meaningful human creative input may still be protected there, as RAND's policy analysis of AI and copyright explains. That single difference changes who controls a commercial image depending on which jurisdiction you rely on.

China has gone further still. In November 2023, the Beijing Internet Court granted copyright to an AI-generated image, finding originality and human intellectual effort in the prompting process — a direct contradiction of the U.S. position that prompts are insufficient. China also mandates labeling of AI-generated content.

The European Union regulates the pipeline rather than the output. Under the EU AI Act, developers must maintain detailed training-data records, and rights holders can object to commercial AI training on their works. For anyone selling images into EU markets, that transparency regime creates obligations that simply do not exist in the U.S.

For businesses and creators working across borders, the practical takeaway is stark:

• A prompt-only image may be uncopyrighted in the U.S. but protected in the UK or China.
• EU market entry means accounting for training-data transparency rules that shape which tools are compliant.
• Protection strategy should follow your primary commercial markets, not your home country's law alone.

There is also a deeper question underneath the jurisdictional map. A Yale Law Journal essay argues that copyright may be the wrong policy tool for generative AI entirely — that expanding it could strain fair use, chill the internet ecosystem, and fail to compensate artists, since litigation power dynamics favor deep-pocketed players. The essay notes that AI copyright cases predominantly involve "big players on both sides," while individual artists find lawsuits expensive, complex, and unpredictable.

For teams building creative workflows around AI images, the divergent rules are one more reason to document human contributions carefully — selection, arrangement, and edits are what matter most under the U.S. framework. Agent skills can help automate that documentation and pipeline work, and SkillzBench is building a free directory where you can browse and test skills before putting them into production.

Your Copyright-Safe AI Image Workflow: 5 Rules to Follow Now

If your team is generating images with AI right now, the rules you follow this quarter could determine whether you own your visual assets — or watch competitors copy them freely. The good news: a defensible workflow isn't complicated. Here are five rules to put in place immediately.

Rule 1: Treat prompt-only outputs as unprotected. Courts have affirmed that human authorship is a "bedrock requirement" of copyright, and the Copyright Office holds that written prompts alone don't give users enough control to qualify as authors. Assume any image generated purely from a prompt can be freely copied by anyone — and plan your branding accordingly.

Rule 2: Build meaningful human editing into every pipeline. AI-assisted works can still be protected when humans select, arrange, or significantly edit outputs — the factors the Office applies case by case. In the Zarya of the Dawn decision, a comic book as a whole earned protection while its individual Midjourney images did not. The dividing line is creative control over the work's expression.

Rule 3: Document your human contributions. As of February 2024, the Copyright Office had registered "well over 100" AI-assisted works — but registration depends on proving which elements humans created. Keep edit logs, version histories, and notes on compositional choices so your claim rests on evidence, not memory.

Rule 4: Disclose AI use in any registration. Applications must name only human authors and include a statement noting AI-generated material. Withholding this information puts the entire registration at risk.

Rule 5: Screen outputs before commercial use. Experiments by Gary Marcus and Reid Southen found Midjourney could generate images that appear to display copyrighted material — a "copyright minefield" for businesses. Before publishing, check for recognizable characters, watermarks, and styles. This matters given that generative AI and copyright are already the subject of more than two dozen lawsuits, including Getty Images' landmark case against Stability AI.

Finally, put a recurring calendar reminder on this landscape. The Copyright Office issued its Part 2 report on copyrightability in January 2025, with training-data guidance following in May — and fair use remains unresolved in pending litigation. What's compliant today may shift within months.

Many of these steps are mechanical: logging edits, running screening checks, maintaining disclosure records. This is exactly where agent skills shine — they can automate image pipelines, edit documentation, and disclosure checklists inside the coding agents and chat-app agents you already use. SkillzBench is building a free directory of vetted agent skills, with security scanning and benchmarking, plus a free MCP connector to wire them into any agent. The waitlist is open now at skillzbench.com — join it, and when the platform launches, your copyright-safe workflow can largely run itself.