AI Generated Content Copyright: What Your Marketing Team Actually Owns

If a model wrote it and you only typed the prompt, you probably do not own a copyright in it. That is the position the US Copyright Office reached in January 2025, after reviewing more than 10,000 public comments. For AI generated content copyright, the line is human control over expression, and your marketing team can stand on the right side of it if it keeps records.

The direct answer. Can AI content be copyrighted? Purely AI-generated material cannot be protected under current US law, and prompts alone do not count as authorship. Human contributions can be protected when they are perceptible in the final work: your own text or images fed in as inputs, your creative selection and arrangement of outputs, and your substantive edits. Document all three.

What the Copyright Office AI report actually concluded

The report is formally titled Copyright and Artificial Intelligence, Part 2: Copyrightability. It was published on 29 January 2025, the second instalment of a multi-part study the Office began with a Notice of Inquiry in 2023. Part 1 covered digital replicas. Part 3 addresses training data, which is a different question from the one marketing teams usually ask.

The Office summarised its own conclusion in one sentence. The outputs of generative AI can be protected by copyright "only where a human author has determined sufficient expressive elements." Everything else in the report is an attempt to say what "determined" means in practice.

The same release states that using AI to assist creation, or including AI-generated material in a larger human-made work, does not bar copyrightability. That sentence matters more to your team than the headline. Most marketing output is assisted, not autonomous.

Prompts are instructions, not authorship

The core finding is about control. The Office concluded that, given current generally available technology, prompts alone do not provide sufficient human control to make the user the author of the output. Prompts, in its framing, function as instructions that convey unprotectable ideas.

The reasoning is practical rather than philosophical. A model fills gaps the prompt leaves open, follows some instructions and ignores others, and returns different results for the same input. The person writing the prompt is closer to a client briefing a commissioned artist than to the artist.

Iteration does not change the analysis. Revising a prompt 20 times gives you 20 more outputs to choose between. It does not, in the Office's view, give you more control over any one of them. Commentators have described this as treating repeat prompting like re-rolling dice.

Three ways a human contribution still counts

The report identifies three routes to protection, and each maps onto ordinary marketing work.

First, expressive inputs. If you feed in your own drawing, photograph or draft and that material is perceptible in the output, your authorship in it survives. Second, selection and arrangement. Choosing which outputs to use and composing them into a larger work can be original authorship, even when the pieces are not. Third, modification. Creative edits to an output are yours, to the extent of the edit.

The Office also declined to recommend new legislation or a special category of rights for machine output. My read is that this is the correct call for now. A new right for unedited output would reward volume, and volume is already the cheapest thing in marketing.

Four decisions that drew the line on AI generated content copyright

The report did not appear from nowhere. It codified a run of registration decisions and one appellate ruling. Read together, they show the boundary more clearly than the report's own prose.

The first case is the one marketers should read closest. Kristina Kashtanova wrote the text of Zarya of the Dawn and arranged it with images generated in Midjourney. In its February 2023 letter, the Office kept protection for the text and for the selection and arrangement of words and images. It refused to treat her as author of the images themselves, and replaced her 2022 certificate with a narrower one.

That is almost exactly the shape of a modern content team's output. A human-written article with AI illustrations, or an ad with human copy over a generated background, is a Zarya-style work. You own the parts you made and the way you put them together.

Thaler is the outer boundary. Stephen Thaler listed his system, the Creativity Machine, as the sole author. Judge Patricia Millett, writing for a unanimous panel, held that the Copyright Act requires a human author, noting that many of its provisions make sense only if an author is a human being. The Supreme Court declined to hear the appeal in March 2026.

Thaler's practical reach is narrow, because no marketing team lists a model as author. Its value is that the human authorship requirement is no longer an agency preference. An appellate court has written it down, and the Supreme Court left it there.

What AI generated content copyright means for blog posts, ads and images

Marketing teams produce three kinds of asset with AI, and the rules land differently on each. The useful question is not "did we use AI?" but "where in this asset is the human expression?"

Blog posts and long-form copy

A post drafted by a model from a one-line brief and published with light proofreading is the weakest case. The structure, sentences and word choices were determined by the system. You can publish it. A competitor who copies it wholesale may face little legal consequence, because there may be little protectable authorship to infringe.

A post where a writer outlines the argument, writes key sections, rewrites the generated passages and chooses the examples is a different object. The human-written and heavily rewritten passages are protected. So is the overall selection and arrangement if it reflects real editorial choices.

My position is blunt. If your content programme depends on long-form posts as a defensible asset, light-touch generation is a poor strategy on copyright grounds alone, separate from quality. The argument about whether volume publishing works at all is covered in the economics of the AI content flood.

Ad creative, images and short copy

Images are where the Office has been strictest, because image models offer the least granular control. A product visual generated from a prompt and used as-is is, on the current reading, not protected. Your competitor could reuse it, and your own registration would have to exclude it.

Images become protectable when you add human authorship the Office can see. Compositing a generated background with your own product photography, painting over generated elements, or arranging several outputs into a designed layout all add a protectable layer. The American Cheese registration shows how far iterative editing can carry you, and also where it stops: the Office protected the arrangement and excluded the individual AI elements.

Short copy is a separate case altogether. Taglines, headlines and slogans were rarely copyrightable even before AI, because the Office has long refused protection for words and short phrases. If your brand line matters, your tool is trademark, not copyright, whether a human or a model wrote it.

Ownership without copyright: what your contracts still control

Most model providers' terms assign output rights to the customer. That is useful, and it is easy to misread. A vendor can assign you whatever rights it has in an output. It cannot assign a copyright that never came into existence, because no human authored the expression.

So "we own our AI content" is often true in the contractual sense and false in the copyright sense. You can use it, the vendor will not claim it, and you can stop your own staff and agencies from treating it as theirs. You may not be able to stop an unrelated third party from copying it. The differences between vendor terms on ownership and indemnity are set out in the comparison of model provider terms.

Agency contracts need the same scrutiny. If an agency delivers AI-assisted work and the contract promises an assignment of copyright, ask what is actually being assigned. I would add a clause requiring the agency to disclose AI use per deliverable and hand over its process records. Without those records you cannot later prove which parts carry human authorship.

Keep a second risk separate in your head. Copyrightability is about what you can protect. Infringement is about whether an output copies someone else's protected work. A purely generated image can be unprotectable and still infringing at once. The infringement side of the ledger is covered in the Disney and Midjourney output case and the training data copyright cases.

How to document human contribution in AI marketing content

The American Cheese registration was refused at first. It was approved after Invoke submitted a recording of the creative workflow. That sequence is the most practical lesson in this whole area: evidence of process converted a refusal into a registration.

Most marketing teams keep no such evidence. Drafts are overwritten in place, prompts live in personal chat histories, and the final file carries no trace of who decided what. If a dispute arises in 18 months, you will be reconstructing authorship from memory.

Keep the inputs and the version history

Save the human-made material that went in: the outline, the writer's draft sections, the source photography, the sketch. Use tools with version history and do not flatten it before archiving. A Google Doc's revision log showing a writer replacing generated paragraphs is better evidence than any affidavit written later.

Store prompts too, but understand their limited value. They show intent and direction. Under the report's reasoning, they do not by themselves establish authorship of the output, so they support your record rather than make it.

Record who selected, arranged and changed what

For each significant asset, a 3-line note in the project tracker is enough. What was generated, what a human wrote or made, and what a human changed. That note is what you would need to fill in the registration form, and it costs a minute at the time versus hours later.

Treat this as the same discipline you would apply to disclosing AI use to customers. The two records overlap heavily, and building one gives you most of the other.

When AI marketing content ownership is worth registering

Copyright exists from creation, so registration is optional. In the US, though, registration is generally required before you sue over a US work, and timely registration affects the remedies available. For most blog posts the cost of registering will not be worth it. For a flagship report, a mascot, a campaign illustration or a course, it often is.

When you do register, disclosure is mandatory. The Office's March 2023 registration guidance requires applicants to disclose AI-generated material that is more than de minimis and to describe the human author's contribution. The AI portion is excluded from the claim.

The cautionary example is again Zarya. Kashtanova's original application did not disclose Midjourney use. Once the Office learned of it, the certificate was cancelled and reissued on narrower terms. A registration that overclaims is fragile, and in litigation the other side will look for exactly that gap.

The Allen dispute shows the opposite failure. The examiner was willing to register Allen's Photoshop edits while excluding the Midjourney material. He declined and asked for the whole image. In its January 2026 cross-motion, the government argued he could likely have registered had he simply disclaimed the material he did not author. My view: claim the human layer, disclose the rest, and do not let an all-or-nothing position leave you with nothing.

Where this argument is weakest

This post leans on the Copyright Office's position as if it settles the matter. It does not, and the gaps are real.

Agency guidance is not a court ruling

The report is the Office's interpretation of existing law. Courts give it weight, but they are not bound by its reading of how much control is enough. Thaler only decided that a non-human cannot be the author. It did not decide whether a human who prompts extensively is one, because Thaler had waived that argument.

That question is precisely what Allen v. Perlmutter puts before a federal court in Colorado. As of the most recent coverage I could verify, summary judgment briefing was complete and no merits decision had been issued. A ruling for Allen would move the line for every marketing team, and the report's prompt analysis would have to be revisited.

The report also qualifies its own conclusion as applying to current generally available technology. Tools that give users fine-grained control over composition, such as region editing, pose controls and reference inputs, push toward the line the Office drew. The analysis is explicitly case by case, so a workflow that fails today may pass in 2 years with no change in the law.

US rules are not global rules

Everything here is US law. Other jurisdictions take different approaches. The UK's Copyright, Designs and Patents Act has long contained a provision for computer-generated works with no human author, assigning authorship to the person who made the arrangements for the work. If your content is exploited mainly outside the US, take local advice rather than generalising from this post.

Finally, for most marketing assets the stakes are modest. Few competitors copy a blog post verbatim, and the remedy for one who does is often a takedown rather than a lawsuit. The documentation habit costs little. The legal exposure from skipping it is, for many teams, small. I still think the habit is worth it, but a team arguing otherwise is not being reckless.

Frequently asked questions

Where to start

Pick the 5 assets your team would be most upset to see copied, whether a pillar report, a campaign visual or a product illustration. For each, write the 3-line authorship note now while people still remember who did what, and attach whatever inputs and version history survive.

Then change one default. Turn on version history in whatever tool your writers and designers draft in, and add an "AI used and where" field to your content brief template. Those two changes cost an afternoon and give every future asset the record that turned the American Cheese refusal into a registration.

Related reading