A Practical Guide to AI-Generated Images and Copyright

The question of AI-generated images and copyright has moved from a curiosity to a routine business decision. Marketing teams generate hero images for landing pages, designers use generative fill to extend backgrounds, agencies mock up campaign concepts in minutes, and e-commerce teams test lifestyle scenes without booking a shoot. Every one of those images eventually raises the same practical questions: who owns it, can anyone stop a competitor from copying it, can it be registered, and what do you owe a client who paid for it?

In the United States the starting point is simple and firm. The U.S. Copyright Office protects only works of human authorship. An image produced by a generative model from a text prompt may therefore not be owned by anyone at all, which is a real problem for a brand that needs exclusive assets and a small problem, or no problem, for a team that just needs a background for a one-week social post. The difference between those two situations is what this guide is about.

This article is written as a decision framework. It sets out the three realistic ways a business can produce imagery when AI is in the picture, the criteria that separate them, a scored comparison, and a recommendation for the situations that come up most often in retouching and design work. It explains how the law currently treats these images as the Copyright Office has described it; it is not legal advice, and anything high-stakes, such as a brand identity or a registration, deserves a conversation with an intellectual property attorney.

What US Copyright Law Actually Says About AI-Generated Images

Four points from the Copyright Office's published position do most of the work in any decision about AI imagery. Everything else in this guide builds on them.

Human authorship is required. Copyright registration in the United States is available only for works created by a human being. A machine cannot be an author, and a work whose expressive elements were determined by a machine does not become protected simply because a person owns the machine or the account.

Prompts alone generally do not make you the author. In its 2025 report on copyrightability, the Office concluded that prompts, on their own, generally do not give the user enough control over the output to make the user the author of the image. The reasoning is practical: the same prompt can produce many different images, and the model, not the person, decides the composition, lighting, rendering and countless other expressive choices. Writing a longer or cleverer prompt does not change that analysis in any predictable way.

Human contributions can still be protected. Where a person selects AI outputs, arranges them into a larger work, or substantially modifies them, those human contributions can be protected. The protection covers what the human did, not the underlying machine-generated material. A retoucher who composites three generated elements with original photography, repaints key areas and builds a new composition may own the composite as a work of authorship, while the raw generated pieces inside it remain unprotected.

Registration requires disclosure. When you apply to register a work, you must identify AI-generated material that is more than minimal and exclude it from your claim. The U.S. Copyright Office's registration guidance for works containing AI-generated material explains how to describe the human-authored portion and disclaim the rest. Failing to disclose is not a paperwork detail; an inaccurate application can put the registration itself at risk.

Three separate questions people tend to merge

Most confusion about AI imagery comes from treating three distinct issues as one. Keep them apart and decisions become much clearer.

• Ownership: Does anyone hold copyright in this image, and if so, in which parts? This is the human authorship question.
• Permission: Does the generation tool's license allow your intended use? This is a contract question, answered by the tool's terms of service, and it applies whether or not the output is copyrightable.
• Infringement risk: Could the image infringe someone else's rights, for example by closely resembling a copyrighted character, reproducing a trademark, or depicting an identifiable person? This is a question about what the image shows, not how it was made.

An image can be unownable, fully permitted under the tool's terms and still risky because it resembles a famous mascot. Another can be partly owned, permitted and low-risk. Your decision needs an answer to all three. For a broader grounding in how licensing works across stock, commissioned and in-house images, our guide to copyright and image licensing covers the non-AI side of the same questions.

The Three Production Routes You Are Choosing Between

In practice, businesses do not choose between "AI" and "no AI." They choose how much of the final image's expressive content comes from a person. That produces three routes worth comparing, each with a different ownership profile.

Route A: Generated and used as-is

A person writes a prompt, perhaps iterates a few dozen times, picks the best output and uses it with only light technical adjustments such as cropping, resizing, format conversion or a minor exposure tweak. The expressive content, meaning composition, subject rendering, lighting and style, comes from the model. Under the Office's current position, this image is very likely not protected by copyright. You can use it, subject to the tool's terms, but so can anyone who finds it.

Route B: AI-assisted, human-led

Generated material is a raw ingredient, not the finished work. A designer or retoucher selects elements from multiple outputs, combines them with original photography or illustration, repaints and reshapes significant areas, and builds a composition that reflects their own creative choices. The human contribution is substantial and visible. The resulting composite can carry protection for the human-authored parts, the selection and arrangement, and the substantial modifications, while the unmodified generated elements inside it remain unprotected and must be disclosed if you register.

Route C: Human-made, with AI tools only for minor edits

The image begins as original photography or illustration, or as licensed or commissioned human work. AI-powered tools may be used for small tasks such as removing a stray power line, extending a background by a few centimeters or cleaning dust spots. The expressive work is human throughout. This route gives the strongest and simplest ownership position, but it costs the most and takes the longest. Where AI edits are more than minimal, the registration disclosure rule still applies, so even this route needs honest records.

Two variations sit at the edges of these routes. Licensed stock that may itself contain AI material falls under whatever the stock license grants, which is a permission, not ownership. And AI used purely for ideation, such as mood boards or rough composition sketches that a photographer then shoots from scratch, leaves the final image fully human-authored, because the generated reference is not in the deliverable at all. That last pattern is one of the most underused options available, and we return to it in the scenarios below.

The Criteria That Decide Between Routes

Six criteria decide most choices. Weight them according to how the image will be used; a single social post and a brand mascot sit at opposite ends of almost every one.

Exclusivity and ownership strength

Can you stop others from copying the image? This matters most for brand identity assets, recurring characters, packaging artwork and anything a client pays a premium to own.

Registrability

Can you register the work, and how much of it? Registration is a prerequisite for filing an infringement suit over a US work and unlocks statutory damages if done in time, so it matters when you expect to enforce your rights.

Speed and volume

How quickly can you produce usable images, and how many variations? Generation wins decisively for concept rounds and high-volume ad testing.

Cost per final image

Total cost, including subscription fees, iteration time, retouching hours, documentation time and review, not just the headline tool price.

Third-party risk

The chance an image resembles a protected character, reproduces a logo or trademark, or depicts a real, identifiable person. This risk exists in all routes but is hardest to control when the model decides what appears.

Documentation burden

How much recordkeeping the route requires to support your ownership claim, satisfy a client, or complete a registration accurately.

There is a seventh criterion that is not legal at all: fitness for purpose. Generated images still struggle in predictable places, including hands, text within the image, product accuracy, consistent characters across a series, and precise color matching to physical products. A route that scores well on cost but produces a product image that misrepresents the item is a poor choice regardless of copyright. For marketplace and catalog work especially, accuracy requirements often decide the matter before ownership is even discussed.

Scored Comparison of the Three Routes

The scorecard below rates each route from 1 (weak) to 5 (strong) against the criteria above. Scores reflect the typical case under the Copyright Office's current position; an individual project can move a point in either direction depending on the tool, the team and the amount of human work involved. For risk and documentation burden, a higher score means less risk or less burden.

Read the scorecard by row, not by total. Adding up the columns suggests the routes are roughly comparable, which hides the real pattern: Route A is excellent on speed and cost and weak on everything related to ownership, Route C is the reverse, and Route B trades a heavier documentation burden for a middle position on almost everything. Your use case tells you which rows carry the weight.

Notice too that Route A scores relatively well on documentation burden. That is not because it needs no records; it still needs a note of the tool, the date and the terms in force. It is because there is little human contribution to document. Route B carries the heaviest burden because its ownership claim depends entirely on proving what the human did.

Route A in Detail: Generated Images Used As-Is

Using generated images directly is fast, cheap and perfectly sensible for a large share of everyday content. The mistake is using it for work that needs to be owned.

Where Route A works well

Short-lived social posts, blog header images, internal presentations, pitch concepts, storyboards and A/B test variations are all reasonable homes for generated images. In each case, the business value lies in having the image now, not in owning it for years. If a competitor copied a generic header image of a coffee cup on a desk, you would lose essentially nothing.

Even here, a few basics apply. Confirm the tool's terms permit commercial use on your plan tier, since some tools restrict commercial use on free tiers or impose conditions. Screen every output for recognizable faces, logos, packaging and characters before it goes live. And keep a simple record so that, if the image later becomes more important than expected, you know how it was made. Our guide to working with AI image generators covers prompting, iteration and quality control for this route in more depth.

Where Route A fails

The classic failure is the generated logo. A founder generates a mark, likes it, puts it on packaging, signage and a website, and only later learns that the artwork itself is very likely unprotected by copyright. Trademark law is a separate regime, based on use of a mark in commerce to identify the source of goods or services, so a generated mark may still function as a trademark; but the business has lost the ability to rely on copyright in the artwork and must depend entirely on trademark rights, whose scope and strength vary. For something as central as a logo, that is a weak foundation, and an attorney should be involved before launch rather than after a dispute.

The second common failure is passing generated images to a client under a standard agreement that promises ownership or exclusivity. If you cannot own the image, you cannot transfer ownership of it. Agencies and freelancers need to disclose AI status and adjust their contracts accordingly.

Route B in Detail: AI-Assisted, Human-Led Images

This is where most professional retouching and design work with AI will land in practice, and where the details matter most. The ownership position is real but partial, and it depends on the human contribution being substantial, deliberate and documented.

What counts as substantial human work

There is no percentage threshold, and anyone who quotes one is guessing. What the Office's position points to is creative control over expressive elements. In retouching terms, these contributions are more likely to count:

• Building a new composition from multiple sources, including original photography, where a person decides placement, scale, perspective and relationships between elements
• Repainting or redrawing significant areas by hand, such as reshaping a figure, rebuilding a product, or replacing a generated face with a photographed model who signed a release
• Creating original lighting and color treatment that changes the character of the image, not just correcting it
• Adding original illustrated or photographed elements that carry real expressive weight

These contributions are less likely to count on their own:

• Cropping, resizing, sharpening and format conversion
• Basic exposure, white balance or contrast correction
• Removing small artifacts or cleaning up extra fingers
• Running the output through another automated filter or upscaler

Technical correction still has great commercial value; it is simply not the kind of work that turns a generated image into a human-authored one. A retoucher who spends two hours fixing hands and matching colors has improved the image enormously without necessarily creating a protectable contribution.

Selection and arrangement

Selection and arrangement deserve special mention because they are the easiest human contribution to overlook. A designer who chooses twelve generated images from several hundred, sequences them into a lookbook and pairs them with original layout and typography may have a protectable compilation even if each individual image is not protected. The protection is thin, since it covers the particular selection and arrangement rather than the images themselves, but it is real, and it is exactly the kind of contribution that must be described accurately in a registration.

Route C in Detail: Human-Made Images With Minor AI Edits

This is conventional photography, illustration and retouching, with AI tools used for small tasks where they save time. It gives the clearest ownership position and is the default recommendation whenever exclusivity is the point.

Keeping minor AI edits minor

The risk in Route C is drift. Generative fill starts by removing a lamp post, then extends the sky, then adds a new building, then replaces the model's jacket. Each step feels small, but by the end a meaningful share of the image may be machine-generated. Set a working rule for your team: removal and small extensions are routine; adding new objects, people, garments or scenery is a decision that gets logged and, for client work, disclosed. A good retouching brief should state which edits are allowed; our guide to writing image briefs for retouchers shows how to put that into a brief so nobody has to guess.

Human-made images also bring their own rights obligations that AI images do not, such as model releases, property releases and licenses for any stock elements composited in. Route C is the strongest on copyright ownership, but it is not automatically clean; it still needs ordinary rights management.

Worked Example: Pricing and Documenting a Mixed Campaign

The following example is illustrative, not a client story. The hours and rates are placeholder assumptions chosen to show how the routes compare; your own figures will depend on tools, team seniority and market rates.

A home goods brand needs imagery for a seasonal launch: one hero image that will run for a year on the homepage, packaging and print catalog; eight product-in-scene images for the online store; and 30 social ad variations for a four-week test. The brand wants to own the hero outright, needs the product images to be accurate, and does not care whether anyone copies the ad variations after the test ends.

Two things stand out. First, documentation is a real line item, roughly 8 percent of the human hours in this example, and it is concentrated in Route B, where the ownership claim depends on it. Second, running all 39 images through Route C would have multiplied shoot and retouching time for images that do not need to be owned, while running the hero through Route A would have saved perhaps a day and left the brand's most visible asset unprotected. Matching the route to the deliverable is where the savings and the protection both come from.

In the product-in-scene images, note why Route B was chosen over Route A. The products must be accurate for the store, since a generated rendering of a lamp that differs from the real lamp creates returns and complaints, so the product itself is photographed. The surroundings, a sunlit shelf or a styled living room corner, can be partly generated because accuracy matters less there. The retoucher's compositing, relighting and color matching are the human contribution. Matching the generated environment to the real product's color is its own craft; see our guide to matching color across images for the method.

Documentation That Supports Your Position

Records are what turn a reasonable ownership claim into a defensible one. They also let you answer a client, a marketplace or a registration examiner quickly and honestly. The goal is not paperwork for its own sake; it is being able to say, for any image, what the machine made and what a person made.

What to record for every AI-assisted image

1. Tool and version: the generator used, the plan or tier, and the date. Terms change, so note the terms version or save a dated copy.
2. Inputs: prompts, reference images and any uploaded source material, including whether you had rights to the references.
3. Raw outputs: keep the unedited generated files, not just the finished image. They show the starting point.
4. Human work: a short plain-language summary of what the person did, such as "composited three generated backgrounds with studio product photo, repainted shelf and shadows, rebuilt lighting."
5. Layered masters: keep layered working files so the human contribution is visible layer by layer.
6. Screening result: confirmation that the image was checked for identifiable people, logos, trademarks and resemblance to known characters.

Store the summary with the file, not in someone's inbox. Embedded metadata can carry creator, rights and AI-disclosure notes alongside the image, and a naming and versioning convention keeps raw outputs and edited masters linked. Our guides to image metadata and rights and versioning image files cover the mechanics. Be aware that metadata is easily stripped by social platforms and some export settings, so treat it as a convenience layer on top of your own records, not a replacement for them.

Registering work that includes AI material

If you register a Route B or Route C image, describe the human-authored contribution in the application and exclude AI-generated material that is more than minimal. A composite might be claimed as "selection, arrangement and modification of AI-generated background elements; original photography of product; digital painting," with the generated backgrounds excluded. Your records make this description easy to write and, just as important, accurate. When the image is commercially important, have an attorney review the application before filing.

Recommendations by Situation

Scores and criteria become useful only when applied. The situations below cover most of the requests an image editing and design team sees. Each ends with a bottom line.

Brand identity: logos, mascots and packaging artwork

These are the assets a business most needs to own, and the ones most likely to be copied if they succeed. Copyright in the artwork supports enforcement against knock-offs, merchandise and lookalike packaging, and trademark protection works alongside it rather than replacing it. Generation can still help at the exploration stage: produce dozens of directions quickly, then have a designer create the final mark from scratch, using the generated options only as reference. The final artwork is then human-authored, and the generated explorations never ship.

Campaign hero and key visuals

A hero image that runs across a website, print and out-of-home for a year or more is a strategic asset. If a competitor lifted it, you would want to act. Route C is the safe default. Route B is defensible when the human contribution is genuinely substantial, for example a composite built around original photography of the product and talent, with generated elements limited to backgrounds or atmosphere and the whole documented. Route A is not appropriate here.

High-volume social ads and short-lived content

For variations that run for days or weeks, ownership matters little, and speed and volume matter a great deal. Route A is a sensible choice, with two guardrails: confirm commercial use is allowed under your plan, and screen every variation for faces, marks and recognizable characters. Platforms also apply their own rules on disclosure of AI content in ads, so check the current policy for each platform you use. Our guide to image editing for social ads covers the sizing and quality side of this workflow.

Product images for stores and marketplaces

Here, accuracy usually decides before copyright does. The product must look like the product. Marketplaces set their own image requirements, and a generated rendering that misrepresents color, proportions or features invites returns and policy problems. Photograph the product and use Route B for scenes and backgrounds, or Route C throughout. Keep AI edits away from the product itself beyond cleanup.

Client deliverables with exclusivity requirements

When a contract promises the client ownership or exclusive rights, you can deliver only what you can transfer. That rules out Route A for anything covered by the exclusivity clause and makes documentation mandatory for Route B. Agree on the AI policy at the brief stage: which deliverables may include generated material, how it will be disclosed, and what the client receives in the handover, such as layered masters, a summary of human work and a list of any generated elements. Build the same questions into client review so approvals are informed; our guide to client proofing for image sets shows how to structure that step.

Editorial and blog illustration

Article illustrations sit in the middle. They are visible and associated with your brand, but rarely worth enforcing. Route A is acceptable for generic illustrations; Route B is worth the extra time for a recurring illustrated style that has become part of your brand's look, since a consistent series is exactly what a competitor might imitate.

Common Mistakes and How to Avoid Them

Most problems with AI imagery do not come from exotic legal edge cases. They come from a handful of predictable mistakes, each easy to prevent once named.

Assuming a generated logo is protected like a designed one

This is the most expensive mistake because it tends to surface late, after the logo is on packaging and signage. Prevent it with a simple rule: final identity artwork is always designed by a person. Generation stays in the sketchbook.

Claiming human authorship for purely generated work

Presenting a generated image as your own authored work, in a registration, a contract or a portfolio, creates problems that are worse than the lack of protection itself. An inaccurate registration can be challenged, and misrepresenting authorship to a client damages trust and may breach the agreement. Describe what you did, precisely, and nothing more.

Using tools whose terms prohibit your commercial use

Generation tools differ on commercial use, attribution, restrictions tied to plan tiers and who bears responsibility if an output infringes. Read the terms before a tool enters production, record which version applied, and recheck when you change plans. The tool's terms govern your permission to use outputs; they cannot grant you copyright that the law does not recognize.

Passing AI images to clients without explaining their status

Clients assume the images they pay for are owned and exclusive unless told otherwise. Tell them. A short AI use statement per project, listing which deliverables contain generated material and what rights attach, prevents disputes and makes your agency easier to trust.

Generating images that imitate identifiable people or trademarks

Prompting for a celebrity lookalike, a competitor's packaging style or a famous character creates risks under publicity, trademark and copyright law that have nothing to do with authorship. Keep names of real people, brands and characters out of prompts and screen outputs for accidental resemblances, which models do produce.

Losing track of which files are which

Months later, nobody remembers which banner was generated and which was shot. Tag AI status in your digital asset management system and folder structure from day one. A single field such as "AI status: none, minor edits, AI-assisted, generated" is enough, provided it is required at upload.

Building an AI Imagery Policy for Your Team

A written policy turns the decisions above into habits. It does not need to be long; one page that everyone actually reads beats a long document nobody opens. A practical policy answers six questions.

1. Which tools are approved, on which plans, and who checked their commercial terms and when.
2. Which asset types may use each route. For example: identity assets Route C only; heroes Route C or documented Route B; social variations any route.
3. What must be recorded for each route, and where those records live.
4. What may never be prompted, such as real people's names, competitor brands, trademarked characters and uploaded images you lack rights to.
5. How AI use is disclosed to clients, on platforms that require it, and in registrations.
6. When to escalate to a senior designer or an attorney.

Review the policy at least twice a year. The Copyright Office has issued guidance and reports in stages and may refine its position as cases and applications develop; generation tools change their terms frequently; and ad platforms and marketplaces update their disclosure rules. Assign someone to watch those sources and flag changes.

When to bring in outside help

Three situations justify professional help. The first is when AI imagery will be central to a brand, for instance a generated visual style that defines the look of a product line, because ownership questions then affect the brand's core value. The second is when registering work that includes AI material, because the application must describe human and machine contributions accurately. The third is when a client requires exclusive rights, because the contract and the production route must match. An intellectual property attorney handles the legal side; a professional retouching and design team handles the production side, building the substantial human contribution and the records that support it. Our image editing and graphic design service works this way on AI-assisted projects, with layered masters and a summary of human work delivered alongside the final files.

Production details that are easy to overlook

A few practical points come up again and again in production and deserve a place in any team's checklist.

Upscaling does not add authorship. Running a small generated image through an AI upscaler to reach print resolution is a technical step. It may be necessary for quality, but it does not change the ownership analysis. Check the upscaled result carefully for invented texture and detail before it goes to print.

Reference uploads carry their own rights. Many tools let you upload an image to guide generation. If that reference is someone else's copyrighted photograph, you may create an output that is substantially similar to it. Upload only images you own or have licensed for that purpose.

Watermarks and credits should be accurate. Adding your studio's copyright notice to a purely generated image implies a claim you probably cannot support. For AI-assisted work, a credit line that reflects the human contribution is both honest and useful. Our guide to watermarks and copyright information covers notice wording and placement.

Consistency across a series is a human job. Models rarely produce the same character, product or environment identically across dozens of images. Maintaining consistency usually requires significant human retouching, which strengthens the human contribution in Route B at the same time as it improves quality.

Keep the raw outputs even when they are ugly. The unedited generations are evidence of what the machine produced and therefore, by contrast, of what the person added. Deleting them to save storage removes the easiest way to show your work.

Taken together, these details reflect one principle. The more an image matters to your business, the more of it should come from a person, and the more carefully you should record what that person did. AI tools are excellent collaborators for speed, exploration and raw material; the ownership, accuracy and accountability still come from human hands.

Where this comes from

• U.S. Copyright Office — Copyright and Artificial Intelligence
• U.S. Copyright Office — Registration guidance for works containing AI-generated material

The figures and practices above come from the sources listed.

Working on something like this?

We take on Image Editing & Retouching work for teams who want it done once, properly. Tell us what you are building and we will tell you honestly whether we are the right studio for it. Start a project.

Where to go next

• Photo Restoration: The Decisions That Matter
• Image Compression and Quality Settings, Done Properly