Can you use AI generated images commercially? Only sometimes. See who owns the output, the infringement risk in training data, and a safe-use checklist.
Nearly every AI-assisted blog ships an AI-generated hero image with every post, this one included, and most never check who owns it or whether the model that made it was trained on someone else's copyrighted art. Can you use AI generated images commercially? Usually, yes. But "usually" is doing real work in that sentence, and it splits into two separate questions that most teams never separate: does the platform's license actually cover commercial use, and could the image itself infringe something a real artist or studio owns. This post answers both, with the current case law and platform terms as of August 2026.
Ownership and permission to use are not the same question, and conflating them is where most of the risk hides. You can have a valid commercial license from your image generator and still not own a copyright in the output, because current US law treats a purely AI-generated image as unowned by anyone, including you.
The US Copyright Office's Part 2 report on copyrightability, released January 2025, concludes that prompts alone don't give a user enough control over the expressive output of a generative model to count as authorship, so an image generated purely from a text prompt can't be registered, per the Copyright Office's own announcement of the report. This isn't a hypothetical position. The Office had already applied it in practice two years earlier: in February 2023 it cancelled part of the copyright registration for the comic Zarya of the Dawn, ruling that its Midjourney-generated images were not the product of human authorship because the creator lacked sufficient control over what the model actually produced, per Cooley's summary of the decision. The comic's human-written text and the human-made selection and arrangement of the images stayed registered. The images themselves didn't.
Register of Copyrights Shira Perlmutter framed the reasoning behind that line directly: "Where that creativity is expressed through the use of AI systems, it continues to enjoy protection. Extending protection to material whose expressive elements are determined by a machine, however, would undermine rather than further the constitutional goals of copyright," per the Copyright Office's own release. The test isn't whether AI touched the image. It's whether a human's choices, not the model's, determined what the final expression looks like.
That same standard is why "AI-generated" isn't a single, binary category for copyright purposes. Hundreds of works incorporating some AI-generated material have been registered since the Office's original March 2023 policy guidance, confirming that mixed human and AI works can qualify when the human contribution is real, according to the same Part 2 report. The line the Office draws in practice: the AI-generated pixels themselves stay unregistrable, but a human's selection, arrangement, and substantial editing on top of them, cropping, compositing, hand-painting over a generated base, meaningfully rearranging elements, can create copyrightable authorship in that layer. A hero banner built by generating a raw image and then hand-adjusting composition, color, and text placement carries a different, stronger claim than a banner used exactly as the model output it.
None of the above stops you from using an image commercially. It just means the license to use it comes from your generator's terms of service, not from owning a copyright in the image. Read those terms by plan, because they vary more than most teams assume. Midjourney's current terms give every paid subscription, Basic, Standard, Pro, and Mega, a commercial usage license, while free and trial access carries only a non-commercial CC BY-NC 4.0 license; separately, any company with more than $1 million in annual gross revenue must be on a Pro or Mega plan to use outputs commercially, regardless of which tier actually generated the image, per terms.law's breakdown of Midjourney's output rights. OpenAI takes a more permissive stance: its current terms assign users full ownership of DALL-E and ChatGPT image outputs and explicitly permit commercial use, marketing, merchandising, resale, on both free and paid tiers, per terms.law's breakdown of OpenAI's output rights. Two of the most-used image generators on the market ship different commercial-use rules by design, which is exactly why checking the actual terms for the plan you're on beats assuming last year's rule still applies.
A valid commercial license from your generator answers the ownership question. It says nothing about whether the image itself infringes someone else's copyright or trademark, and that's a separate, live legal question the courts are still actively deciding.
The first major ruling on generative-AI training landed in the UK, and it split its findings in a way worth understanding precisely. In Getty Images v. Stability AI, decided by the UK High Court on November 4, 2025, the court rejected Getty's copyright claim over Stable Diffusion's model weights, finding no evidence of infringing copying that occurred within the UK during the model's development. But it found Stability liable for trademark infringement on a narrower, separate ground: Stable Diffusion had historically output synthetic images carrying Getty's own watermark, which courts treat as passing off a mark, not as copying an image, per Pinsent Masons' coverage of the ruling. Copyright and trademark are different rights, and this case shows they can point in opposite directions on the exact same set of facts.
The parallel US case is still live. In Andersen v. Stability AI, artist Sarah Andersen's direct copyright-infringement claim against Stability AI is proceeding through discovery, after a judge accepted the theory that copyrighted works can be represented inside a diffusion model as algorithmic or mathematical patterns, rejecting the argument that storage in a different medium defeats an infringement claim. Trial was originally set for September 8, 2026, but the court has since pushed it to April 5, 2027 as discovery in the case expanded, per the AI Lawsuit Tracker's case summary. Whichever way the fair-use question ultimately lands, the case is a live signal that training on scraped copyrighted art carries real exposure for the model provider, exposure that sits upstream of anything a user does with the output.
The case that matters more directly for anyone prompting a generator, rather than building one, is the one Hollywood filed against Midjourney. On June 11, 2025, Disney, NBCUniversal and DreamWorks sued Midjourney in the US District Court for the Central District of California, alleging the tool could be prompted to output near-exact copies of copyrighted characters, per Georgetown Law's Tech Institute summary of the filing. The complaint cites Midjourney outputs carrying the likeness of Yoda, Marvel superheroes, characters from Aladdin, the Minions, Bart Simpson and Shrek, with side-by-side comparisons against the originals built into the filing itself, per Decrypt's coverage of the suit. The studios asked the court to bar Midjourney from operating without real infringement safeguards in place.
This is the case that separates two risks people tend to collapse into one. Whether training a model on copyrighted images is fair use is still an open legal question, the whole subject of the Andersen case above. Whether a specific output, prompted to reproduce a specific named character, infringes that character's copyright is a much narrower and much more answerable question, and it doesn't require resolving the training question first. You don't need a court to settle model training to know that prompting for "Bart Simpson at the beach" and publishing the result on a commercial blog is a bad idea.
The Getty ruling above is the clean illustration of a risk that has nothing to do with your prompt's creativity: a generator can leak artifacts of its training data straight into your output, visibly. Getty won its trademark claim specifically because Stable Diffusion's images sometimes carried Getty's own watermark baked into the pixels, proof the model had memorized parts of specific training images rather than synthesizing something new. A hero banner that quietly ships with a distorted logo, a faint signature, or a stock-photo watermark in the corner isn't a hypothetical edge case. It's the exact fact pattern a UK court already ruled on. It's also the single easiest thing on this list to catch, because it doesn't require a lawyer, just a look at the finished image before it publishes.
None of the risk above requires legal review on every banner. It requires four checks, run once per image before it ships, the same discipline this blog applies to every factual claim in a post's text.
Pull up your actual current plan's terms before assuming last year's understanding still holds. As of August 2026: Midjourney grants commercial rights on every paid tier, Basic through Mega, but requires Pro or Mega specifically once your company crosses $1 million in annual revenue, per terms.law. OpenAI grants full ownership and commercial rights on both free and paid ChatGPT and API tiers, per the same source. Neither of those terms has stayed static over the past few years, and neither will necessarily stay static going forward, so this is a check to repeat, not a fact to memorize once.
The Disney and Universal suit against Midjourney exists because a prompt naming a specific character is the single most reliable way to get a near-exact, clearly infringing reproduction out of an image model. The same logic applies to prompting in a specific living artist's style, which is a separate legal theory (a Lanham Act false-association claim) running alongside the copyright claims in the Andersen case. A blog banner that needs "a coral glass shield" doesn't need "in the style of [named illustrator]" to get there. Write prompts around concepts and visual descriptions, not proper nouns tied to someone else's copyrighted work.
Before a banner ships, look at it the way Getty's own testing looked at Stable Diffusion's outputs: is there a faint watermark, a distorted signature, or a logo-like mark baked into the image. Run a reverse image search on anything that looks even slightly like it could be lifted from a specific stock photo or a named artist's portfolio. This is a five-minute look, not a legal review, and it's the same instinct behind this blog's own fact-checking discipline applied to pixels instead of prose: a fluent, confident output isn't the same thing as a verified one.
Indemnification, where the platform itself agrees to cover your legal costs if an image gets challenged, is rare and worth knowing you don't have it by default. Adobe offers IP indemnification to commercial users of Firefly, a policy it can make because Firefly trains on licensed Adobe Stock and public-domain content instead of scraped web images, per Computerworld's coverage of the offering. Shutterstock sells indemnification against copyright and trademark claims on AI-generated images, but only to Enterprise customers who route the image through a human review step first, per its own announcement; tellingly, Shutterstock won't even accept AI-generated submissions from its own individual contributors, because it can't verify clean IP ownership on their behalf either, per Shutterstock's contributor content policy. Midjourney and OpenAI, the two tools most SaaS blogs actually reach for, offer no indemnification at all: the commercial license is real, but the infringement risk stays with you.
This is a different exposure than the two other legal checks this blog has covered for AI-assisted content. The EU AI Act's Article 50 asks whether you have to disclose that a post is AI-generated; it's a transparency duty about your text, not a question of who owns your images or whether they infringe anything. Trademark risk in comparison posts covers a different exposure again: naming a competitor's brand in the body of a post, which has its own doctrine (nominative fair use) that has nothing to do with image copyright. And Google's ranking treatment of AI content is a fourth, unrelated axis entirely: Ahrefs' analysis of 600,000 pages found essentially no correlation between a page's AI-content share and where it ranks, so a copyright-clean AI banner and a copyright-risky one rank identically in Google's eyes. None of these four checks substitute for each other. A blog running an AI writer needs all of them, and the practical way most teams end up running all four consistently is putting the review where the draft already lives: a pull request, where an image and its prompt sit next to the diff a reviewer is already reading before merge, the same review surface an AI blog writer built for developers uses for the fact-check pass on the text.
A checklist only works if someone actually runs it on every post. Lyra opens each post as a pull request with its banner and sourcing attached to the same diff a reviewer already checks before merge, so a risky prompt or a lookalike image gets caught there, not after it ships.
FAQ
Usually yes, but two separate questions decide it: whether the platform's terms grant you a commercial license, and whether the image itself infringes something a real artist or studio owns. A paid Midjourney or ChatGPT plan generally licenses commercial use. Neither company can guarantee the output doesn't resemble copyrighted training data closely enough to create liability, which is a separate risk you carry, not one the license waives.
Under current US Copyright Office guidance, nobody holds a copyright in an image generated purely from a text prompt, because a prompt doesn't give the user enough control over the expressive output to count as authorship. The image sits in a legal gray zone: usable, but not ownable, and not something you can stop someone else from copying. Real, substantial human editing on top of the raw output, compositing, hand-drawn additions, meaningful arrangement, can bring the human-authored portions back under copyright.
It's possible if the image closely resembles a specific artist's style, a studio's copyrighted character, or a real photograph the model trained on, and the rights holder notices. The risk is highest with named-character prompts (Disney, Marvel, Star Wars, and similar), stock-photo lookalikes with a faint embedded watermark, and outputs that closely mimic a living artist's signature style. Generic, non-referential prompts carry far less exposure.
No, not as a standard feature. Midjourney and OpenAI's terms grant commercial usage rights but explicitly leave infringement risk with the user. Indemnification, where the vendor agrees to cover your legal costs, exists only from a small number of providers and usually only on enterprise tiers: Adobe covers commercial Firefly users because Firefly trains on licensed Adobe Stock and public domain content, and Shutterstock offers it to Enterprise customers only, built around a human review step before the image ships.
Built by the tool you're reading about
Lyra finds the topics worth ranking for, writes them in your repo's voice, fact-checks every claim, and opens a pull request scored and ready to merge. You review and hit merge. Want to see what she'd write for you? Start free with three posts, no card.
Keep reading
Multi-agent content review pairs a generator with an independent critic and judge so mistakes get caught before a human ever opens the pull request.
AI content governance means proving who wrote a post, who reviewed it, what changed, and when it shipped, an audit trail a Git pull request already gives you.
AI blogs sound generic without a real guide. What belongs in a brand voice AI content style guide, and how to keep it current as your product evolves.