The Image No One Can Own
In February 2023, the U.S. Copyright Office made a decision that sent shockwaves through the creative technology world. It partially revoked copyright protection for Zarya of the Dawn, a graphic novel created by artist Kris Kashtanova using Midjourney, an AI image generator. The written text and the arrangement of images? Those Kashtanova could keep. The AI-generated illustrations themselves? Those belonged to no one.
“The images in the Work that were generated by the Midjourney technology are not the product of human authorship,” the Copyright Office ruled in a letter that has since become a landmark document in the emerging field of AI intellectual property law.
It was a seemingly narrow administrative decision. But it cracked open a fault line that runs beneath the entire generative AI economy — one that implicates billion-dollar companies, working artists, legal scholars, and anyone who has ever typed a prompt into an AI tool and clicked “generate.” The question it raised wasn’t merely technical or legal. It was philosophical: what does it mean to create something, and who deserves to benefit from that creation?
No one, it turns out, has a fully satisfying answer. Not yet.
A Technology That Outpaced the Law
To understand why the ownership question is so vexed, you need to understand how generative AI art actually works — and how radically different it is from every creative technology that came before it.
Tools like Midjourney, DALL-E 3, Stable Diffusion, and Adobe Firefly are built on diffusion models — complex neural networks trained on billions of image-text pairs scraped from the internet. These models learn statistical relationships between visual patterns and descriptive language. When a user types “a lonely lighthouse at dusk, oil painting style,” the model doesn’t retrieve an existing image or even consciously compose one. It reconstructs a plausible visual output from learned noise patterns, guided by that text prompt.
The output can be stunning. It can look indistinguishable from work produced by a skilled human artist. And it can be generated in roughly four seconds.
This speed and accessibility have produced staggering volume. Adobe’s Firefly model alone generated over 3 billion images in its first three months of availability in 2023. Midjourney, operating primarily through Discord, has attracted more than 16 million users. Estimates from market researchers at Cognitive Market Research suggest the AI art generation market will reach $1.8 billion by 2030.
Copyright law, meanwhile, was designed for a world where authorship was self-evident. The Statute of Anne, enacted in England in 1710, protected authors of books. The first U.S. Copyright Act of 1790 assumed human creators. Every subsequent revision — the landmark Copyright Act of 1976, the Digital Millennium Copyright Act of 1998 — operated on the same foundational premise: a human being made this thing.
Generative AI broke that premise entirely.
“Copyright law has a human authorship requirement that’s been there from the very beginning, and it’s actually quite well-established in the case law,” says Ryan Abbott, a professor of law and health sciences at the University of Surrey and one of the leading scholars on AI and intellectual property. “The question is whether there’s any role for human creative expression in the process, and how much of that is sufficient.”
The Three-Sided Ownership Argument
When legal scholars, artists, and technologists debate who should own AI-generated images, three broad positions have emerged — each with legitimate claims, none fully adequate.
The user owns it. This is the position most intuitive to anyone who has spent hours crafting, refining, and iterating on AI prompts. Prompt engineering, its advocates argue, is a genuine creative skill. Getting a useful output from Midjourney or DALL-E requires specific knowledge of how to structure language, what stylistic references to invoke, and how to guide the model through multiple iterations. Jason Allen, who won first place in the digital art category at the Colorado State Fair in 2022 with an AI-generated piece called Théâtre D’Opéra Spatial, spent approximately 80 hours refining his prompts in Midjourney before arriving at his winning image. That feels like authorship to many people.
The Copyright Office has shown some openness to this view in narrow circumstances. When a human selects, arranges, or meaningfully modifies AI outputs, the resulting compilation or modification may be protectable. But a prompt alone, the office has consistently said, is insufficient to establish authorship over the generated image itself. The gap between imagining an outcome and controlling the precise execution of it is simply too large.
The AI company owns it. If the tool is doing the creative work, perhaps the company that built and trained the tool deserves the intellectual property rights. Several AI companies have tried to claim ownership through their terms of service, then quietly reversed course. OpenAI’s early DALL-E terms assigned ownership to the company; later versions transferred rights to users while granting OpenAI a broad license. Midjourney’s terms give users ownership of outputs for personal use but retain rights for commercial applications unless users pay for higher-tier subscriptions.
Legal experts are nearly unanimous that AI companies cannot claim copyright on generated outputs under current law, precisely because copyright requires human authorship — and a corporate entity that didn’t create the work through human expression doesn’t qualify. The more interesting, and alarming, question is what AI companies do own: the models themselves, the training pipelines, and the proprietary systems that generate value from other people’s creative work.
The artists whose work trained the model own it. This is the position advanced by a growing coalition of working artists, and it may be the most morally compelling argument in the debate — even if it’s the hardest to enforce legally. Generative AI models are trained on enormous datasets that contain copyrighted images, often scraped without permission or compensation. Artists including Sarah Andersen, Kelly McKernan, and Karla Ortiz filed a class-action lawsuit in January 2023 against Stability AI, Midjourney, and DeviantArt, alleging that the companies had “scraped” billions of copyrighted images from the internet to build their commercial products without consent or payment.
A separate lawsuit filed by Getty Images against Stability AI in both the U.S. and the UK alleges that the company copied more than 12 million images from Getty’s library without authorization. Getty’s UK lawsuit notably included the observation that some AI-generated images contain faint, mangled versions of the Getty watermark — a ghostly reminder of the images that were consumed in the training process.
As of mid-2024, none of these cases had reached final judgment, but the courts’ eventual decisions will be seismic.
What the Courts Are Actually Deciding
The legal landscape shifted meaningfully in August 2023, when a federal judge in Washington, D.C. upheld the Copyright Office’s refusal to register an AI-generated artwork called A Recent Entrance to Paradise, created autonomously by a system called DABUS, developed by computer scientist Stephen Thaler. U.S. District Judge Beryl Howell was unambiguous: “Human authorship is a bedrock requirement of copyright.”
But the more nuanced and commercially consequential cases involve the training data question. In February 2024, Judge William Orrick in the Northern District of California allowed portions of the artists’ class-action lawsuit against Stability AI to proceed, while dismissing several claims. The court found that direct copyright infringement claims related to training data needed more factual development, but kept alive claims about the generation of images in the style of specific artists.
The style question is particularly treacherous legal territory. Copyright law in the United States explicitly does not protect style — only specific expression. You cannot copyright “impressionism” or “photorealism.” But the line between style and expression blurs when an AI can generate an image that, to any reasonable observer, is indistinguishable from the work of a specific living artist. When users instruct Midjourney to generate an image “in the style of Greg Rutkowski” — a Polish digital artist whose name became one of the most commonly used prompts in early AI image generation — the resulting images compete directly with Rutkowski in commercial markets, even if no single image is technically copied from him.
“My name became a prompt,” Rutkowski told MIT Technology Review in 2022. “People are copying my style to get jobs that would otherwise come to me.” He later joined the efforts to have his work removed from Stable Diffusion’s training dataset.
The European Union has moved faster than the United States in attempting to regulate this space. The EU AI Act, which passed in 2024, requires providers of general-purpose AI models to publish “sufficiently detailed summaries” of the content used to train their systems — a transparency requirement designed in part to help copyright holders identify whether their work was used without consent. Whether that transparency requirement will translate into meaningful enforcement or compensation remains to be seen.
The Creative Economy at Stake
The stakes of the ownership question extend far beyond legal theory. The answer will determine whether working artists can survive in an AI-saturated creative economy.
The economic disruption is already measurable. Stock photo agencies have seen significant shifts in demand patterns, with Shutterstock reporting in 2023 that it was generating tens of thousands of AI images daily through its own licensed AI tools — while simultaneously facing competition from unlicensed AI generators. Getty Images, by contrast, has pursued a different strategy: licensing its library to AI companies under controlled terms, launching its own AI generator in 2023, and aggressively litigating against those who didn’t pay.
The impact on individual artists has been more severe and less evenly distributed. Illustrators, concept artists, and stock image creators — whose work is highly reproducible and highly prompt-able — have seen commission rates and demand drop noticeably in some markets. A 2023 survey by the Animation Guild found that 44 percent of animation professionals believed AI had already affected their employment prospects. Video game concept artists have reported project cancellations and budget reductions attributed to AI substitution.
Defenders of AI art tools argue that the technology democratizes creativity, allowing people without formal artistic training to produce visual work that communicates their ideas. They also contend that artistic progress has always involved building on prior work — that human artists absorb influence, study masters, and learn styles without compensating those who came before. Is training an AI model fundamentally different from the way a human art student fills sketchbooks copying Rembrandt?
“There’s a legitimate philosophical debate there,” acknowledges Andres Guadamuz, a reader in intellectual property law at the University of Sussex. “But there’s also a practical economic argument that’s separate from the philosophical one. Whatever we decide philosophically, we need a system that allows creators to sustain themselves economically. That’s what copyright was designed to do.”
A Path Forward — Or Several Competing Paths
The most thoughtful observers of the AI copyright debate tend to agree that existing law is insufficient and that new frameworks are necessary — but they disagree sharply on what those frameworks should look like.
One emerging proposal is a system of compulsory licensing, similar to the mechanical license that governs music. Under such a scheme, AI companies would be required to pay into a fund when using copyrighted works for training, and that fund would be distributed to artists based on some measure of how heavily their work was used. The music industry established a rough precedent with the Sound Exchange model for digital performance royalties, which collects and distributes payments at scale without requiring individual negotiation.
Another approach, championed by some AI companies, is opt-out registries — databases maintained by AI companies or neutral third parties that allow artists to request their work be excluded from training data. Stability AI, Spawning AI, and others have experimented with these systems, but critics point out that opt-out places the burden on artists who never consented in the first place, and that enforcement is essentially impossible to verify.
A third camp argues that AI companies should simply face liability for training on copyrighted content without licenses, forcing the market to find a solution. If Stability AI faces a billion-dollar judgment in the Getty case, the argument goes, the economics of unlicensed training data become untenable, and licensed data partnerships — like those Getty and Shutterstock have already struck — become the industry norm.
What seems increasingly clear is that the current ambiguity serves primarily one constituency: the AI companies themselves. When ownership is unclear, enforcement is impossible, accountability is deferred, and the economic benefits flow to those who have already captured them. The artists, users, and legal system are left sorting through the wreckage of a property regime that no one bothered to update before deploying a technology at civilizational scale.
The U.S. Copyright Office released a significant policy report in early 2024 outlining a range of issues it was actively studying, signaling that new guidance on AI-generated works was coming. Congress has held multiple hearings on AI and copyright, with legislators from both parties expressing concern — though actual legislation remains elusive.
Judge Howell’s ruling that “human authorship is a bedrock requirement” will not be the last word. It may not even survive the appellate process as more nuanced cases wind through the courts. The law, as it always does, will eventually catch up to the technology. The question is how much damage accumulates in the gap — and whose creative livelihood and legacy it takes with it.
The image generated in four seconds may not belong to anyone right now. But the fight over who gets to say that — and what happens next — is the defining intellectual property battle of our time.