Who Owns an AI-Generated Image? Copyright, and the Firefly Loophole

Ask a lawyer whether an AI-generated image can be copyrighted and you'll get a careful, hedged answer. Ask a marketing team whether they can use one in a campaign and you'll get a shrug. The gap between those two responses is where a lot of money is being spent in 2026, and a lot of risk is being quietly carried.

The state of the law, briefly

The core question is whether a work produced mostly by a machine can have a human author for copyright purposes. The US Copyright Office has been fairly clear that purely AI-generated output doesn't qualify. The murkier territory is the middle ground: what happens when a human writes a prompt, iterates on the result, edits it, and composes it into something larger.

That middle ground is still unresolved. The practical effect is that the copyright status of unedited AI output remains uncertain everywhere, and any claim to ownership rests on how much human authorship went into the final image, not on the generation itself.

What this means in practice: you can generate an image, but you may not be able to register a copyright on it, and enforcing a claim against someone who copies it is an uphill fight.

The licensing side is a separate mess

Copyright is about what you can claim. Licensing is about what the platform lets you do, and the platforms have set very different terms.

OpenAI assigns output rights to the user. Google and the other major API providers generally do the same, with acceptable-use restrictions on top. Midjourney takes a different approach: companies above $1 million in annual revenue have to be on the Pro or Mega plans to use outputs commercially, and Midjourney offers no IP indemnification.

Indemnification is the part that matters for businesses. It's a promise that if someone sues you for copyright infringement over a generated image, the platform will cover the legal cost. Most platforms don't offer it.

The Firefly exception

Adobe Firefly is the outlier. It's the only major generator that offers full indemnification, and it's built that way for a specific reason: Adobe trained it on content it has the rights to, including licensed stock imagery and public domain work.

That training-data strategy changes the risk profile. A model trained on scraped web content inherits some uncertainty about where its outputs came from. A model trained on licensed data can make a cleaner claim, and Adobe backs it with a legal promise.

The tradeoff is style. Firefly is competent but not the aesthetic champion. Midjourney wins the art-direction conversations; Firefly wins the procurement conversations. For an agency or a brand that can't afford a copyright claim, the safe-but-less-exciting option is often the right one.

The training data question

Underneath the licensing terms is a deeper issue that won't go away quietly: what the models were trained on.

A model trained on scraped web content carries uncertainty about whether its outputs borrow too closely from a specific artist's work. Courts are still working through the cases, and the answers will shape what "AI image ownership" means for years. For now, the practical effect is that a closed model's training data is a black box, and no one can guarantee a generated image is free of influence from copyrighted work.

Adobe's approach was to sidestep this by training Firefly on content it licenses, including stock imagery and public domain material. That doesn't make the model more creative. It makes the legal position cleaner, and Adobe pairs it with indemnification.

Open-source models add another wrinkle. A model like Stable Diffusion or Flux can be fine-tuned on almost anything, which means the responsibility for what went into it shifts partly onto whoever did the fine-tuning. That flexibility is the point of open source, but it also means there's no single vendor to point at when something goes wrong.

A practical checklist

If you're about to use AI images in something that matters, run through four questions before you publish.

One, who made the final image? The more human editing went into it, the stronger your claim to authorship. Keep the process documented.

Two, what does the platform's license say? Check whether commercial use is allowed at your revenue level, and whether the platform indemnifies you against infringement claims.

Three, what was the model trained on? If you can't answer this, assume some risk and price it accordingly.

Four, does the project justify the risk? A personal blog post and a national ad campaign are not the same thing. Match your caution to the stakes.

What this means for you

If you're an individual making images for yourself, the risk is low. Generate, edit, and don't worry much about registering a copyright you'll probably never need to enforce.

If you're making images for a client or a company, the calculus changes. The moment an image goes into a paid campaign, you're exposed to whatever the training data and licensing terms inherited. That's when indemnification stops being a legal nicety and becomes a purchasing requirement.

The honest position in 2026 is that the law hasn't caught up to the tools. Copyright law was written for authors and artists, not for prompts and diffusion models. Until it does, the practical answer to "who owns this" is: you can use it, you probably can't fully protect it, and if you're spending real money on it, read the licensing terms before you publish anything.

It's worth restating the asymmetry, because it drives most of the anxiety. Generating an image is trivial. Owning it, protecting it, and defending it are not. The gap between those two is where the entire copyright conversation around AI image generation lives, and no model release is going to close it. Only the courts will, and they're moving slowly.

The safest path, for now, is the one that requires the least faith in unresolved law: do real human work on top of the generation, keep your editing process documented, and pick a platform whose indemnification matches the stakes of the project.