Who Owns an AI Image? The New Copyright Battles in Digital Art

Why training data, licensing, and authorship are becoming the central legal questions of our creative era.

An illustrator I know discovered last year that her work was in an AI training dataset. Not because anyone asked permission — she found out when someone showed her an AI-generated image that had her distinctive watercolor texture, her specific way of rendering light, even compositional choices she’d developed over fifteen years. The AI hadn’t copied any single piece exactly. It had learned from her work, then generated something new but unmistakably influenced by it.

She asked me what her legal options were. I didn’t have good answers.

Copyright law was designed for a world where copying was clear — you either reproduced someone’s work or you didn’t. AI complicates this fundamentally. The systems learn from copyrighted works without reproducing them exactly, then generate outputs that might resemble the training data to varying degrees. Traditional legal frameworks struggle with this, creating uncertainty about ownership, compensation, and rights that affects everyone from individual artists to major corporations.

Scraped Without Permission: How Training Datasets Are Built

Most major AI image generators were trained on datasets built by scraping billions of images from the internet. The companies argue this constitutes fair use — they’re not distributing the copyrighted works, just learning from them the way a human art student might study existing work. Artists argue this is unauthorized commercial use of their intellectual property at massive scale.

The scale matters. A human art student might study hundreds or thousands of works over years of training. AI models ingest millions of images in weeks, processing them into mathematical patterns that can then generate unlimited new images. The companies building these systems become commercially valuable partly through that training — Midjourney, for instance, is valued at over $1 billion, built substantially on training data that artists never consented to providing.

Several major lawsuits are working through courts now. Getty Images sued Stability AI in 2023, claiming the company used millions of copyrighted images from Getty’s catalog without permission. A class action by visual artists including Sarah Andersen, Kelly McKernan, and Karla Ortiz makes similar claims against Stability AI, Midjourney, and DeviantArt. These cases will likely establish precedents that shape the industry.

When an Output Resembles a Copyrighted Work

Even if courts eventually allow AI training on copyrighted work, there’s a separate question: what happens when AI output closely resembles specific copyrighted pieces?

Early versions of Stable Diffusion could be prompted to generate images in the style of specific living artists with remarkable accuracy. Some outputs were similar enough to raise questions about whether they constituted derivative works — which would violate copyright if created without permission. More recent model versions include some safeguards, but the fundamental capability remains.

A photographer I spoke with — we were at a conference in Austin, conversation happening in the hallway between sessions — described discovering AI-generated images that mimicked his signature techniques so closely that people assumed they were his work. “I spent years developing that style,” he said, visible frustration in how he gestured. “Now anyone can generate something that looks like mine in thirty seconds. How is that not taking something that belongs to me?”

The legal question is whether style can be copyrighted. Generally, US law says no — copyright protects specific expression, not general style or technique. But AI might push courts to reconsider. When technology makes style perfectly replicable, does that change the analysis?

Who Deserves Credit: Prompt Writer, Developer, or Original Artist?

Assuming AI-generated images can be owned at all — and that’s not certain — who owns them? The person who wrote the prompt? The company that built the model? The artists whose work trained it?

Most AI companies’ terms of service grant users rights to outputs they generate, with some restrictions. Midjourney, for example, gives paid subscribers ownership of their generations. But this only addresses the relationship between company and user — it doesn’t resolve questions about the artists whose work contributed to training.

There’s also genuine ambiguity about creative contribution. If someone spends hours crafting prompts, iterating through hundreds of generations, curating and selecting outputs, are they creating something? Or are they just good at using a tool that does the actual creative work? The answer might matter legally — copyright generally requires human creative input.

(Though honestly, I’m not sure where the line is. Is writing a prompt fundamentally different from directing a photographer or giving detailed instructions to a commissioned artist? Maybe. Maybe not. The law will probably land somewhere messy and compromised.)

The U.S. Copyright Office and the Human Authorship Requirement

In 2023, the US Copyright Office issued guidance stating that AI-generated content lacks human authorship and therefore cannot be copyrighted. The decision came after several test cases, including an attempt to register a fully AI-generated graphic novel.

The guidance allows for copyright of works that combine AI-generated elements with sufficient human creative input — selection, arrangement, modification. But purely AI-generated images enter the public domain immediately upon creation. Anyone can use them, reproduce them, modify them without permission.

This creates interesting situations. If you generate an AI image and use it commercially, competitors can legally copy it because you don’t own it. This might push commercial users toward hybrid approaches — using AI as a starting point but adding enough human modification to claim copyright.

Other jurisdictions are taking different approaches. The UK and EU are developing frameworks that might grant some rights to AI outputs. China has indicated AI-generated content might be copyrightable under certain conditions. This jurisdictional variation creates complexity for anyone working internationally.

Artists Demand Transparency: What Was My Art Used For?

Many artists aren’t necessarily opposed to AI training on their work — they’re opposed to it happening without knowledge, consent, or compensation. This has sparked demands for transparency about what’s in training datasets.

Some tools now let artists check if their work is in specific datasets. Have I Been Trained, created by Spawning AI, allows artists to search for their images in LAION-5B, a massive dataset used to train Stable Diffusion and other models. Many artists discovered their entire portfolios were included without their knowledge.

Knowing is one thing. Having recourse is another. Even after discovering their work in training datasets, most artists lack practical options. They can’t remove work that’s already been used for training. They can’t demand compensation retroactively. They can try to prevent future use, but images already circulating online might get scraped regardless.

A digital artist I know spent days submitting opt-out requests to multiple AI companies after finding her work in several datasets. “It felt like trying to get water back into a bottle,” she told me during a late-night video call, exhaustion evident in her voice. “Even if they honor the opt-outs going forward, the damage is done. My style is in there. It’s learned.”

Toward Fairer Systems: Consent, Licensing, and Compensation Models

What would fair use of artistic work in AI training look like? Several models are emerging, though none has achieved widespread adoption.

Opt-in systems where artists explicitly consent to inclusion in training datasets, possibly in exchange for compensation or access to the resulting tools. Adobe’s Firefly was trained this way — using Adobe Stock images where contributors had agreed to AI training.

Licensing schemes where AI companies pay for training data, similar to how stock photo companies license images. This could provide revenue streams for artists while giving companies legal certainty.

Attribution and compensation mechanisms built into AI systems — technology that tracks which training images influenced specific outputs, enabling automatic credit and payment. Several startups are developing these systems, though implementation at scale remains challenging.

The challenge is economic. Training datasets contain millions of images. Licensing each one individually would be prohibitively expensive. Collective licensing — similar to music rights organizations — might work, but requires infrastructure that doesn’t yet exist.

I keep thinking about that illustrator whose work was in the training data without permission. We talked about it over coffee last month — cold afternoon, the cafe warm and crowded. She’d initially been angry, then resigned, now she’s somewhere between the two. “I don’t know if fair is even possible at this point,” she said. “The technology moved faster than any framework for handling it ethically. Now we’re trying to retrofit fairness onto something that was built without it. That’s probably the best we can do. But it’s not great.”

The copyright questions surrounding AI-generated art are far from resolved. Courts are still working through cases that will set precedents. Copyright offices globally are developing frameworks that might diverge significantly. Artists, companies, and users all have stakes in outcomes that remain genuinely uncertain. What’s clear is that our existing copyright system — built for a world of direct copying — struggles to handle technologies that learn patterns and generate new work. How we adapt these systems will determine who benefits from AI art and who bears its costs.

Questions to consider:

If AI systems can learn from copyrighted work without copying it, does that learning constitute fair use — or does it exploit legal ambiguity to avoid compensating creators?

When AI-generated images can’t be copyrighted, are we creating a two-tier creative economy where human-made work has legal protection and AI-generated work doesn’t?

Is it possible to build fair compensation systems for artists whose work trains AI, or has the technology already moved too fast for equitable solutions?

Jean Marie Bonthous (publishing as JM Bonthous) is the author of two books on digital art and one on AI-created music:

How to Thrive in the Digital Art Market: the Artist’s Guide to Online Platforms, Collectors, and Trends

The Digital Artist’s Guide to Success: How to Create, Promote, and Profit from Your Art

Mastering SUNO: The Ultimate Guide to AI Music Creation

(All available on Amazon)

He has also authored seven other books on the human side of artificial intelligence and six on nonfiction filmmaking. His writing also explores the intersections of AI, digital art, and nonfiction filmmaking. See his latest books: www.jmbonthous.com

He also blogs on MEDIUM, in a separate blog, about the human side of AI: https://medium.com/@jm_26203

Connect with him on MEDIUM and follow https://medium.com/@jmbonthous to stay in the loop with the latest stories about digital art and with @jmbonthous1 to stay in the loop with the latest stories about the human side of AI.