Who Owns the Copyright of AI-Generated Images?

Correction (19 October 2026): We corrected our account of Naruto v. Slater (the Ninth Circuit ruled that animals lack standing to sue; the human-authorship rule comes from the US Copyright Office and Thaler v. Perlmutter), of the Zarya of the Dawn decision (the registration was narrowed, not revoked), and of Turkish law (there is no Yargıtay ruling on AI output yet, and unprotected output is not an "anonymous work").

Today, someone with no formal art training can produce striking images in the style of Rembrandt, or hyper-realistic photographs, within minutes by typing a short text prompt. Who owns the result? The person who wrote the prompt? The company that built the algorithm? The owners of the images the algorithm was trained on? Or is it an ownerless heap of data sitting in the "public domain"?

This article examines a problem that goes to the foundations of intellectual property law. Drawing on historical precedents, recent decisions of the US Copyright Office (USCO), the European Union's new AI Act, and Turkey's Law on Intellectual and Artistic Works (FSEK), it considers the question under three headings: "Background," "Implications," and "Conclusions."

Can an algorithm's random arrangement of pixels express an artistic will?

1. Background: The Anatomy of Copyright and Historical Precedents

To understand the problem AI poses, we first need to understand why modern copyright law exists. Intellectual property law was created to reward "human intellect and creativity" and to encourage social progress by giving artists economic protection.

1.1. "Sweat of the Brow" vs. "Modicum of Creativity"

Historically, two doctrines have competed over when the law should protect a work. Under the "Sweat of the Brow" doctrine, which originated in English law, the labor and time spent creating a work (e.g., alphabetizing a massive telephone directory) were enough to earn copyright. The US Supreme Court rejected this doctrine in its well-known Feist Publications v. Rural Telephone Service (1991) decision. The Court held that effort alone is not enough; a work must possess at least a "Modicum of Creativity." Alphabetizing a phone book was a mechanical task, not a creative choice.

This is precisely where today's debate over AI prompts sits. Is typing the instruction "draw a red horse running under a blue sky" a spark of creativity, or a mechanical instruction, much like telling a taxi driver, "take me to Times Square"?

1.2. The Monkey Selfie Case (Naruto v. Slater)

Before the AI lawsuits, the strangest "non-human creator" case the law had faced involved a macaque named Naruto. Nature photographer David Slater left his camera in the jungle, and Naruto used it to take a selfie. The photo went viral. PETA (People for the Ethical Treatment of Animals) sued Slater, claiming the copyright belonged to the monkey.

This decision became the starting point for the AI debate: nothing can be copyrighted without human will and individual creative input.

2. The Global Legal Framework and Recent Rulings

Although national legal systems differ slightly on who owns AI-generated images, a broad international consensus has begun to emerge.

2.1. US Copyright Office (USCO) and the "Zarya of the Dawn" Case

The most important milestone in this field is artist Kris Kashtanova's application to the US Copyright Office for the comic book *Zarya of the Dawn*. Kashtanova wrote the text of the comic book themselves but generated all the visuals using Midjourney (AI). The USCO initially granted registration but, on learning that the images were AI-generated, cancelled the original certificate and replaced it with a narrower registration covering the text and the selection and arrangement of the images, excluding the Midjourney images themselves.

The USCO's Ruling: The comic book's texts, page layout, and the way the visuals were curated are copyrightable because they contain Kashtanova's human creativity. However, the individual images produced by Midjourney hold no copyright. The USCO likened a user entering a prompt to a client giving instructions to a painter. The client can say, "use red paint, draw a tree here," but the painter (i.e., the algorithm) decides where the brushstrokes go and how the light falls. The outcome cannot be predicted: between the image in the author's mind and the image the AI produces lies a gap filled by the machine's own random choices. Purely AI-generated output therefore cannot be copyrighted.

2.2. Turkey: Law No. 5846 on Intellectual and Artistic Works (FSEK)

Article 1/B of FSEK, the core of Turkish copyright law, sets out the basic condition for a product to qualify as a "work" (eser): "Any kind of intellectual and artistic product bearing the characteristic of its author..."

In Turkish law, "characteristic" (hususiyet) means that the personality, intellectual effort, spirit, and creative choices of the author are reflected in the work. An algorithm has no soul or personality, so a machine's output cannot bear a "characteristic" in this sense. There is as yet no ruling of the Yargıtay (Court of Cassation) on AI output, but applying its general case law on hususiyet, which requires a human author, the prevailing view in Turkish doctrine is that an image produced entirely by giving prompts to an AI (with no human modification) is not a work under FSEK. Within Turkey, such images are therefore unprotected (free to use). They are not "anonymous works": under FSEK that term covers protected human works whose author is not named.

3. "Fair Use" and the Hidden Theft Crisis: Training Data

Looking only at the "output" side leaves the picture incomplete. The biggest fight is on the "input" side. How do Midjourney or Stable Diffusion know Van Gogh's brushstrokes, or the style of a well-known living illustrator? Because these models are trained on millions of copyrighted images "scraped" from the internet without permission (the LAION-5B dataset is one example).

Getty Images sued Stability AI, alleging that it had used millions of Getty's licensed photographs without permission to train its model. Distorted versions of the Getty Images watermark were even visible in some outputs. AI companies argue that this is "machine reading," comparable to a person visiting a library, looking at thousands of pictures, and "taking inspiration" to draw their own, and so falls under "Fair Use." Authors reply that this is not "inspiration" but the direct copying of works into model weights to build a "commercial imitation machine."

4. Implications: How Will Industries Transform?

This legal uncertainty, and the absence of copyright, are having a major impact on the creative industries:

1. Advertising and Marketing Agencies: Agencies used to pay tens of thousands of dollars to stock photo sites (Shutterstock, etc.) for a campaign poster. Today they can generate such images with AI in seconds, at no cost. The catch is that they do not own the copyright. If a rival copies the AI-generated image exactly and uses it in its own advertising, the agency has no basis to object (at least as regards the image). For brands, this is a serious corporate-identity risk.
2. Freelance Illustrators: The jobs of digital artists, especially those drawing backgrounds or creating character concepts in the gaming industry, are directly threatened. Companies prefer AI output that is "good enough," even if it cannot be copyrighted, to paying for human labor. The democratization of art (anyone can now produce images) brings with it the devaluation of art.
3. Film and Gaming Sectors: Studios such as Marvel or Disney do not want AI-generated characters or scenes, because a character they cannot copyright and license (or sell toys of) brings them no billion-dollar return. Large studios are therefore pushed to invest in fully human-made work, which copyright protects.

5. Conclusions: What Awaits Us in the Future?

The current copyright regime (the human-centered rules of FSEK or the USCO) is unlikely to hold up for long against such fast-moving technology. In a world where 90% of artworks are produced by machines in seconds and immediately fall into the public domain, the internet would become a vast "uncopyrighted visual wasteland." What, then, is the solution?

Legal scholars and policymakers are debating three main scenarios:

• The Zero Protection Scenario: The status quo continues and machine output remains in the public domain. This could deter companies that want copyright protection from using AI.
• Return to "Sweat of the Brow": If a person spent days refining a prompt, tested 50 different words, and "curated" the result from hundreds of variations, that labor itself could be recognized as a new kind of "Work."
• Creation of a Sui Generis Right: AI-generated works would not receive full "Copyright" (70 years of protection is too long); instead, a new status, an "AI Output Generator Right" lasting 5 or 10 years, would be created along the lines of the neighboring rights enjoyed by music producers, solely to prevent commercial copying.