AI-Generated Images for Commercial Use: What the Terms and Copyright Law Actually Say

AI-Generated Images for Commercial Use: What the Terms and Copyright Law Actually Say

An AI image generator hands you a finished picture in seconds and a rights question that takes longer to answer than the render did. The download button says nothing about who owns the picture, whether a client can use it in an ad, or whether anyone can stop a competitor from using a near-identical image next week. Those answers live in three different places: the platform terms you accepted, copyright law in your country, and the rights of people and brands that appear inside the picture.

This guide reads those three layers separately, using the terms and help pages published by OpenAI, Adobe, and Leonardo AI, and the published position of the US Copyright Office on AI-generated material. It covers what each platform grants, where the free plans change the deal, why a commercial-use grant is not the same thing as copyright ownership, and a checklist to run before an AI image ships in client work, a product, or a paid campaign.

This is practical guidance for working designers, not legal advice. If a project carries real legal exposure, take the terms named here to a lawyer.

The Short Answer

Three questions get collapsed into one, and they have different answers.

Did the platform give you permission to use the image commercially? Usually yes. OpenAI assigns its rights in the output to you. Adobe allows commercial use of Firefly output from features that are not labelled beta, and allows most beta output commercially too, with a catch covered below. Leonardo AI allows commercial use on both free and paid plans. Permission, though, is only the first layer.

Do you own copyright in the image? Often not, or not fully. The US Copyright Office position, set out in its January 2025 report on copyrightability, is that copyright protects human authorship, and that prompts alone do not give the user enough control over the expressive elements of an output for the user to count as its author. Human contributions can be protected: material a person authored that remains visible in the output, a sufficiently creative selection or arrangement of generated material, or human modifications that themselves clear the originality bar. A raw output with no such contribution may have no copyright owner at all, which means you can use it and so, in copyright terms, can anyone else who gets hold of it.

Does the image clear everyone else's rights? No platform grant answers this. If an output resembles a real person, carries something close to a brand's logo, or looks substantially like an existing copyrighted work, the problem belongs to you. OpenAI's terms make the user responsible for their content. Adobe's generative AI guidelines prohibit prompts designed to produce copyrighted or trademarked output. The platform gave you a tool and a permission. It did not insure the picture.

The Decision Table

Read the table as a floor, not a conclusion. Every row still has to pass the copyright question and the third-party rights question before an image is safe to ship.

OpenAI: You Own the Output, and So Might Someone Similar

OpenAI's Terms of Use, effective January 1, 2026, cover ChatGPT and DALL-E. The ownership clause is short and generous: as between you and OpenAI, you retain ownership of your input and you own the output, and OpenAI assigns its right, title, and interest in the output to you. There is no split between free and paid accounts in that clause, and no attribution requirement attached to it.

The next clause is the one designers skip. Under the heading "Similarity of content," the terms state that output may not be unique and that other users may receive similar output, and that OpenAI's assignment does not extend to other users' output. Two studios can prompt their way to near-identical hero images, each holding a valid assignment from OpenAI, with neither able to claim the other copied them. For campaign work, that rules out treating a raw AI image as an exclusive brand asset. It works fine as an illustration. It fails as the one picture your client's identity depends on.

Three further duties in the same terms matter in production:

• You are responsible for the content. You warrant that you hold the rights needed for anything you feed in as input, and you carry responsibility for making sure your use does not break the law or the terms.
• You may not present output as human-made when it was not. The terms put this in the list of prohibited uses. Credit lines, case studies, and contest entries that describe an AI image as a photograph or a hand-drawn illustration breach it.
• You may not use output to build a competing model. Fine for design work, relevant if a client wants generated images stockpiled as training data.

OpenAI also offers no indemnity in these consumer terms and gives no promise about what the model was trained on. The grant is broad and simple, and the risk stays with the user.

Adobe Firefly: Built for Commercial Use, With a Beta Line

Adobe takes the opposite approach to OpenAI on training data, and says so plainly. Firefly's first commercial model was trained on licensed content, including Adobe Stock, and on public domain content where copyright has expired. That provenance is the core of Adobe's commercial-safety pitch: the inputs were licensed, so the outputs start from a cleaner position than models trained on unlicensed scraping.

Adobe's Firefly FAQ divides output by feature status, and the division is worth memorising:

• Features without the beta label: outputs can be used in commercial projects.
• Beta features: outputs can be used commercially unless the product or other Adobe material designates otherwise, but those outputs are not eligible for indemnification while the feature stays in beta.

Indemnification is the part enterprise clients ask about, because it decides who pays if a third party claims an image infringes their work. A beta badge quietly removes that cover while leaving commercial permission in place. Before a large campaign, record which feature generated each image and whether that feature carried a beta label at the time. The design asset licensing checklist gives a place to file that record alongside the rest of the project assets.

Adobe's Generative AI User Guidelines, updated May 15, 2026, add user-side duties that sit on top of the licence:

• Using the tools to create content that violates a third party's copyright, trademark, privacy, or publicity rights is prohibited. The guidelines name prompts designed to generate copyrighted or trademarked content, and reference images you upload that contain someone else's work.
• Impersonating other people or entities is prohibited, and so is disseminating misleading or deceptive content that could cause real-world harm.
• Adobe may attach Content Credentials to content made or modified with its generative features, as a signal that AI was involved. You must not remove, alter, or disable those credentials. Stripping the metadata to make an image pass as a photograph is a terms breach, not a workflow trick.

That last point changes delivery habits. A pipeline that strips all metadata from exports, which many teams run automatically, can remove Content Credentials along with the camera data. Check what your export preset actually removes before Firefly images go through it.

Leonardo AI: The Plan Decides Who Owns the Picture

Leonardo AI is the clearest example of a plan changing ownership rather than features. Its Terms of Service and its Commercial Usage help article draw the line in the same place:

• Paid subscribers generating privately hold full ownership of their images, including copyright as between them and Leonardo.
• Free users can also use their generations commercially. Leonardo's help centre says this in direct terms, with a T-shirt business as its example. But under the terms, ownership of a free user's output vests in Leonardo on creation, and free generations are public.

Public is the operative word. For public content, Leonardo takes a broad, perpetual licence to use, reproduce, modify, publish, and distribute the content, including for training models and for commercial purposes, and other users can access and reuse public assets through the service as its features allow. So a free-plan image on a client's product packaging is permitted, and it is also available for a competitor to find, reuse, and sell on their own packaging. Leonardo's own commercial-use article makes exactly this point with its two examples: the free user and the paid user can both sell their designs, but the free user's design may be used by others too.

The terms add the same uniqueness warning as OpenAI's: outputs may not be unique, and your ownership does not extend to other users' outputs.

For client work on Leonardo, the working rule is simple. Generate on a paid plan, set generations to private, and keep them private. The price of the plan is buying exclusivity and privacy, not buying commercial permission, which the free plan already includes.

The Copyright Gap No Platform Term Can Close

Platform terms are contracts between you and the platform. Copyright is a separate system, run by national law, and a contract cannot create a copyright that the law does not recognise.

The US Copyright Office set out its position in the second part of its Copyright and Artificial Intelligence report, published in January 2025. The position, in plain terms:

• Copyright requires human authorship. AI-generated material, as such, is not protected.
• Using AI does not automatically disqualify a work. Human-authored expression that remains perceptible in the final piece can be protected.
• A human who selects or arranges AI-generated material in a sufficiently creative way can hold copyright in that selection or arrangement. A human who modifies generated material enough that the modifications themselves are original can hold copyright in those modifications.
• Prompts alone do not provide sufficient control over the expressive elements of the output for the prompter to be treated as the author of those elements. A detailed prompt is instructions to a system that makes the expressive choices, closer to commissioning than to drawing.

The Office had already applied this logic in practice. In the Zarya of the Dawn case, it kept registration for a graphic novel's human-written text and for the arrangement of the work, and cancelled protection for the individual AI-generated images inside it. Designers outside the US should treat the details as American law, but the underlying pattern, protection follows the human contribution, is the direction several copyright systems are taking. Local advice decides local cases.

What this means at a desk, in production terms:

Logos and brand marks are the weakest use. A raw AI logo may be commercially permitted by the platform and still be unprotectable as copyright, hard to defend if copied, and risky if it drifts near an existing mark. Our free fonts for logos guide covers the type side of the same problem. If a mark matters, a human designer should rebuild and refine it into original vector work, which creates human authorship and a cleaner trademark story.

Illustration and campaign imagery are the middle ground. Permission is clear on all three platforms. Protection comes from what your team adds: composition across a layout, retouching, combination with photography and type, and selection from many candidates. The finished advertisement as a whole can carry protected human authorship even where a single generated element does not.

Merch and resale designs carry the exclusivity problem. Print-on-demand and stock resale depend on stopping copycats. A design that anyone can legally reuse, because it sits in a public free-plan gallery or because no copyright subsists in it, cannot do that job. Sell AI-based merch only where you hold ownership under a paid private plan and have added enough human work to defend the design. Compare the resale limits in our free illustrations guide, where the same pattern appears in conventional licences.

Contracts need the disclosure conversation early. If a client is paying for original artwork, delivering raw AI output without saying so risks breaching both the brief and, on OpenAI, the term against presenting output as human-generated. Put the tool, the plan, and the rights position in the project note before delivery, not in an email after the client asks.

The Traps, Named Plainly

The trademark in the picture. Generators produce plausible logos, labels, cars, and product shapes unprompted. A generated background with a readable, near-real brand mark inside it is a trademark problem the platform licence does not solve. Zoom in and read every sign, shirt, and badge in the frame before shipping, the same habit our free stock photos guide recommends for photographs.

The face that looks like someone. A generated person who resembles a real, identifiable individual carries publicity and privacy risk, and presenting that face as endorsing a product adds a false-endorsement problem. Adobe's guidelines prohibit impersonation outright. Treat any output face that looks like a specific person, celebrity or not, as unusable for advertising.

The style prompt aimed at a living artist. Asking for work in the style of a named living illustrator, or tied to a current film or game franchise, aims the tool at someone else's rights. Adobe prohibits prompts designed to generate copyrighted or trademarked content. A style is not always protected by itself, but the output of such a prompt is the highest-risk output a generator produces, and the client file will show the prompt if a dispute starts.

The beta badge you did not notice. On Firefly, a beta label leaves commercial permission in place and removes indemnification. Feature status changes over a project's life, so record it at generation time rather than reconstructing it later.

The public gallery on a free plan. On Leonardo, free generations are public by design. Client concepts, unreleased product renders, and campaign art do not belong there. This is a confidentiality failure before it is a licensing one.

The metadata strip. Automated export pipelines that remove all metadata can remove Content Credentials from Adobe-generated images, which Adobe's guidelines forbid. Test the pipeline, not just the image.

The uniqueness assumption. OpenAI and Leonardo both warn in their terms that outputs may not be unique. Never promise a client that no similar image exists, and never build a trademark filing on the assumption that a generated design is one of a kind.

A Worked Example: Clearing AI Images for a Small Campaign

The following is a hypothetical example, written to show the checking order. No real client, product, or figures are involved.

A designer is preparing a launch campaign for a small homeware brand: one hero image for the website, three supporting illustrations for product pages, a background texture for packaging, and a logo refresh.

The audit runs per image, in this order:

1. Hero image, generated in Firefly through a released feature. Adobe's FAQ allows commercial use of non-beta output, and the training provenance is licensed and public domain material. Verdict: clear for the website. The feature name and its non-beta status are written into the project note, and the export preset is checked so Content Credentials survive.
2. Supporting illustrations, generated on a free Leonardo account. Commercial use is allowed, but the images are public and ownership vests in Leonardo. A competitor could reuse the same illustrations on rival product pages. Verdict: not cleared for a brand campaign. The set is regenerated on a paid plan with private generations, and the free-plan versions are discarded.
3. Packaging background, generated with ChatGPT. OpenAI assigns its rights in the output to the user, so permission is clear. A close look at the texture shows a woven label shape carrying lettering that reads like a real textile brand. Verdict: regenerate with the label area prompted out, then re-check the new output at full size.
4. Logo refresh, generated output used as a sketch. The raw output is not shipped. A designer redraws it as original vector work, changing proportions and details in the process. Verdict: the redrawn mark, not the render, goes to the client, with the AI sketch kept in the file as a concept reference only.

Total time: about an hour, most of it spent on the logo decision and the Leonardo regeneration. The campaign ships with every image tied to a named platform, plan, and feature status, and the two riskiest assets, the public illustrations and the logo, are the ones that got rebuilt.

The Pre-Delivery Checklist

Run this before any AI-generated image leaves your machine. It extends the pre-shipment pass in the design asset licensing checklist with the AI-specific steps.

• Platform, plan, and feature recorded: every image has a note naming the generator, the plan in force, and, for Firefly, whether the feature was beta at generation time.
• Ownership checked against the plan: private paid generation where exclusivity matters; no client work sitting in a public free-plan gallery.
• Human contribution identified: for any image the client needs to protect, the file shows what your team added: selection, arrangement, retouching, redrawing, or layout.
• Trademark pass at full size: every logo, label, sign, and product shape in the frame read and cleared, or the image regenerated.
• People pass: no generated face resembles a real identifiable person, and no person in the image appears to endorse the product.
• Prompt record kept: prompts saved with the project, with no prompt aimed at a living artist's style or a protected franchise.
• Credentials and metadata intact: export settings checked so Content Credentials on Adobe-generated images are not stripped.
• Disclosure made: the client note states which images are AI-generated and what rights position they carry. No image is described as a photograph or hand-drawn work.
• No exclusivity promised beyond the evidence: the client hears, in writing, that similar outputs may exist unless the work was privately generated and substantially reworked.

A 60-Second Decision Tree

For any AI-generated image you are about to ship, answer in order. Stop at the first no.

1. Can you name the platform, plan, and feature that made it? If the file is a mystery download, it is not cleared for client work.
2. Does that combination allow commercial use, and who owns the output under it? Free plans can allow the use and keep the ownership. Know which deal you hold.
3. Is the image private and exclusive enough for this job? Campaigns, packaging, and merch need private generation. A public gallery image fails here.
4. What did a human add? If the answer is only a prompt, expect no copyright in the image itself. Add real selection, arrangement, or modification where protection matters, or accept the weaker position in writing.
5. Does anything inside the frame belong to someone else? Logos, labels, faces, and franchise imagery each fail the image on their own. Regenerate or rebuild.
6. Will the delivery survive an audit? Prompts, plan records, feature status, and the disclosure note are in the project file. If not, finish the paperwork before the image ships.

FAQ

Can I use AI-generated images in client work?

Yes, on the platforms compared here, provided the plan and feature allow commercial use and the image passes the rights checks in this guide. OpenAI assigns its rights in output to the user on free and paid accounts. Adobe allows commercial use of Firefly output from non-beta features, and of most beta output without indemnification. Leonardo allows commercial use on free and paid plans, with ownership and privacy split by plan. Client contracts can still restrict AI use, so check the brief and disclose the tool before delivery.

Do I own the copyright in an image I generate?

Not automatically. Under the US Copyright Office position published in January 2025, prompts alone do not make the user the author of the output, so a raw output may carry no copyright at all. Human contributions can be protected: your own authored material visible in the piece, a sufficiently creative selection or arrangement of generated elements, and original modifications you make. Platform terms decide ownership between you and the platform, which is a contract question, separate from whether copyright exists.

Can I use an AI-generated image as a logo?

It is permitted by the platforms and weak in every other respect. A raw generated logo may have no copyright protection, cannot promise uniqueness, and may sit close to an existing mark. Use generation for concepts, then have a designer rebuild the chosen direction as original vector artwork. The rebuilt mark carries human authorship and a far cleaner position for trademark review.

Can I sell AI-generated designs on shirts and other merch?

Only where the deal you hold supports resale and some level of exclusivity. A free Leonardo design is public, owned by Leonardo under its terms, and reusable by others, so it cannot anchor a product line. A privately generated image on a paid plan, substantially reworked by you, is the stronger base. Marketplace and print-on-demand platforms also set their own rules on AI content and disclosure, which sit on top of the generator's terms.

Is Adobe Firefly safer than other generators for commercial work?

Adobe builds its case on training data: its first commercial Firefly model was trained on licensed content such as Adobe Stock and on public domain content, which reduces the chance that an output reproduces unlicensed work. Its FAQ also ties indemnification to feature status, with beta outputs excluded while in beta. That is a cleaner starting position, not a guarantee. The trademark, likeness, and prompt checks still apply to every image.

Do I have to tell anyone an image was AI-generated?

Often, yes, in one form or another. OpenAI's terms prohibit representing output as human-generated when it was not. Adobe attaches Content Credentials to Firefly images and prohibits removing them. Client contracts, advertising rules, and marketplace policies can each add their own disclosure duties. The safe default is to disclose in the project paperwork and keep credentials intact, even where no label appears on the image itself.

Another site is using an image almost identical to mine. What can I do?

Start by checking which position you actually hold. If your image was a raw output, the platforms warn that similar outputs go to other users, and copyright may not subsist in the image to enforce. If you privately generated the base and added substantial original work, your position is stronger, and the project records, prompts, drafts, and layered files are the evidence. That is a conversation for a lawyer, armed with the file history this guide recommends keeping.

What to Read Next

• Free stock photos for commercial use - the photography counterpart, where model and property releases join the file licence.
• Design asset licensing checklist - the pre-shipment pass across fonts, icons, photos, mockups, and AI images in one place.
• Free illustrations for commercial use - conventional illustration licences compared, including the resale limits AI plans echo.
• Free icons for commercial use - MIT, ISC, Apache 2.0, and CC0 sets, and what each licence asks of you.
• Image formats guide - choosing the delivery format once an image is cleared to ship.

This guide is maintained by the Lucky Graphics editorial team and reviewed when the platforms above change their terms. It is general information about how AI image terms and copyright basics interact, not legal advice for a specific project.