15 min read
The US Copyright Office asked an image generator for a bespectacled cat in a robe, reading a newspaper. Its detailed prompt specified fog, lighting, composition and a lens. The output added a human hand nobody had requested. The example in the Office’s January 2025 report exposes the authorship question: how much of the final expression did the person control?
A business publishing generated content needs to establish its ownership. If a competitor copies it, copyright protection depends on the original human expression in the work and the applicable jurisdiction. Paying for an AI service or spending hours prompting it does not by itself settle that question.
The U.S. report distinguishes purely generated material from human work made with AI assistance. A publisher needs to identify the contribution it can claim before treating the whole output as an exclusive asset.
What the Copyright Office Ruled
In January 2025 the US Copyright Office published Copyrightability report. It is the second in a series; it followed a notice of inquiry that drew more than 10,000 comments from all 50 states and 67 countries, and roughly half of those comments addressed exactly this question.
Four of the report’s eight conclusions apply to anyone commissioning work
- Using AI tools to assist rather than stand in for human creativity does not affect the copyright protection available for the output.
- Copyright protects the original human expression in a work even where the work also contains AI-generated material.
- Copyright does not extend to purely AI-generated material, or to material where there is insufficient human control over the expressive elements.
- On current generally available technology, prompts alone do not provide sufficient control.
The fourth point can come as a surprise, because it holds regardless of the effort put into the prompt. The Office addresses the effort argument directly and rejects it, pointing to Feist Publications v. Rural Telephone Service, the 1991 Supreme Court decision establishing that copyright protects original authorship rather than hard work. The report says revising a prompt repeatedly is like re-rolling the dice, generating more outputs to choose from without changing your degree of control over any of them.
The cat example tests that control directly. The system followed some instructions, ignored others and supplied expressive details itself. A detailed request can influence the result without making the person the author of everything that appears in it.
This is not just an agency view. In Thaler v. Perlmutter, decided 18 March 2025, the DC Circuit held that the Copyright Act requires all eligible work to be authored in the first instance by a human being. The Supreme Court declined to hear an appeal on 2 March 2026, docket number 25-449, which leaves that ruling standing. The case concerned a work claimed to be entirely machine-authored; it does not decide every mixed human–AI workflow.
Read only the headline and you would think AI assisted work is uncopyrightable. The report reaches a more specific conclusion. It identifies three specific human contributions that do produce protectable expression.
If you feed in something you created, an illustration, a photograph, a piece of writing, and that work is still visible in what comes out, you hold copyright in that part. The Office describes the scope as analogous to a derivative work, your expression is protected, the AI generated elements around it are not.
Assemble AI generated material into a whole with enough creative judgement and the original compilation may be protectable, even though the components individually are not. One practical limit from the Office’s Copyright compendium, a selection of only two or three elements is generally treated as too minimal to qualify.
Original creative edits may carry protection for those changes. The scope is case-specific; editing does not automatically extend rights to every generated element. The report specifically notes that tools letting users select and regenerate individual regions of an image, rather than reprompting the whole thing, can amount to genuine control over placement of creative elements in a way that plain prompting does not.
Sitting underneath all three is the assistive use carve out, which is broader than most people realise. The Office lists colour correction, detail sharpening, de-blurring, song ideation and generating a preliminary outline for a piece of writing as uses that do not affect copyright in the finished work at all. If you brainstorm with a model and then write the thing yourself, nothing about your copyright changes.
Abstract tests are hard to apply. Two actual Copyright Office decisions make the line concrete, and they went opposite ways for a specific reason. A comic book with Midjourney generated images and human written text. In its February 2023 decision, the Office protected the text and the selection, coordination and arrangement of words and images across the pages. It refused protection for the images themselves. Copyright protects the book without protecting the generated pictures themselves.
An applicant uploaded their photograph and instructed a system to render it in the style of Van Gogh’s The Starry Night. The Office’s Review Board upheld the registration refusal, finding insufficient creative control. Compare that with a work called Rose Enigma, where an artist’s hand drawn illustration remained clearly perceptible in the output. Rose Enigma was registered, with the registration annotated to cover only the human authorship visible in the deposit.
The difference is not effort or intent. It is whether a human authored work survives visibly into the output. The input was absorbed in one case and remained perceptible in the other.
One more thing that catches people out. The Office’s registration guidance requires applicants to disclose AI generated content that is more than de minimis, and to describe the human contribution. This is a disclosure obligation on you, not a question the examiner will guess at. The Office has registered hundreds of works containing AI generated material on exactly that basis.
The Unsettled UK Rule
If you sell into the UK, the position is not the same, and this is where a lot of generic advice written for a US audience stops being accurate.
Section 9(3) of the UK copyright law covers literary, dramatic, musical and artistic works generated by computer in circumstances where there is no human author. It assigns authorship to the person by whom the arrangements necessary for the creation of the work are undertaken. Protection runs for 50 years from creation under section 12(7).
The provision predates generative AI by roughly three decades, and no reported UK case has applied it to an AI output. Ireland, New Zealand, Hong Kong and India carry comparable provisions, though their wording and case law need separate analysis. This article does not establish equivalent protection across those jurisdictions.
It may not last. The UK government ran a copyright and AI consultation from December 2024 to February 2025 that drew over 11,500 responses. Under sections 135 to 137 of the Data (Use and Access) Act 2025, it was required to report to Parliament by 18 March 2026. The report landed on the deadline, and as Bird analysis, removing section 9(3) protection was one of the few firm outcomes in it. The primary report proposes continued monitoring and removal absent evidence of ongoing value. A policy proposal does not itself repeal section 9(3).
If you are relying on section 9(3) to claim ownership of AI output in the UK, you are relying on a provision whose application to generative outputs and future policy remain uncertain. The uncertainty leaves a thin foundation for an asset you are counting on.
EU Disclosure Rules
Ownership is one question. Disclosure is a separate one, with different duties for providers and deployers. EU AI Act, Regulation (EU) 2024/1689, has Article 50 transparency obligations applicable from 2 August 2026. It imposes four transparency duties, and it is not limited to high risk systems. Providers of systems that interact directly with people must make clear a person is dealing with AI. Providers of generative AI systems have duties to mark covered synthetic outputs in a machine readable format, detectable as artificially generated. Deployers must label deepfakes and AI generated text published on topics of public interest.
Two details affect commercial use. First, the Article 50(4) labelling duty for AI generated text does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication. Whether the exception applies depends on those conditions; a nominal review label is not enough. Second, the European Commission published the first draft of the labelling code in December 2025, followed by a final voluntary code on 10 June 2026, and machine-readable marking is a provider obligation, distinct from deployer labelling. The Commission’s current explanation describes a grace period until December 2026 for the marking duty on generative systems placed on the market before 2 August. The exceptions and dates must be checked for the particular system and use.
Penalties under the Act reach 15 million euro or 3 percent of worldwide annual turnover, subject to the applicable penalty provisions and proportionality rules, including the distinct treatment of smaller businesses. These are maximums, not an automatic charge for every unlabelled output.
Three Practical Cases
Before the formal audit steps below, here is a faster way to see whether this applies to you. Read each scenario and check whether it matches something you shipped this year.
Case one. A marketing team generates twenty product description variants from a single prompt, picks the one that reads best, and publishes it unedited. In the U.S., the unchanged generated text alone is unlikely to establish human authorship. Picking a favourite from a batch is what the Office compared to re-rolling dice, since the control was exercised over which output to keep, not over the expression inside it.
Case two. A designer sketches a rough layout by hand, feeds it to an image generator as a reference, and the generator’s output keeps the original composition recognisable. The original human expression that remains perceptible may be protected. This is the Rose Enigma pattern. The original expression has to still be visible in the deposit copy, not just present in your process.
Case three. An editorial team runs fifty AI generated article drafts through a real edit, rewriting structure, adding original analysis, and cutting anything unverifiable. The team’s original analysis and creative changes may be protected; the full set of generated material does not automatically become protected. Substantial creative modification is one of the three contributions the Office recognises outright, and the editorial trail itself becomes the evidence if anyone ever asks you to prove it.
Assess the contribution asset by asset. Neither the number of drafts nor the existence of an editorial process proves authorship of every expressive element.
Documenting Human Authorship
Nobody asks you to prove authorship on a good day. They ask during an infringement dispute, a registration application, or an acquisition where someone is valuing your content library. All three arrive after the work is done, which is exactly when reconstructing what happened is hardest.
Four records can make that assessment easier. They document the work; they do not guarantee protection or resolve a disputed claim.
The useful question is not whether AI was used. It is which of the three protected contributions is present. Label each asset as assistive use, human input perceptible in output, or selection and arrangement plus modification. Under the U.S. framework, a lack of original human expression weakens a copyright claim in generated material, and you should know which assets those are before you build a strategy around them.
Rose Enigma succeeded because the hand drawn illustration was still perceptible in the final image, and the applicant could show it. SURYAST failed because the input did not survive visibly. Archive the source illustration, draft, photograph or recording alongside the finished asset. Storage is cheap. Reconstructing an input you deleted eight months ago is not possible.
The Office has been explicit that prompt volume and prompt effort do not establish authorship. What does establish it is selection, arrangement and modification, and those are decisions, not inputs. A short note per asset covering what was chosen from what, how it was arranged, and what was changed afterwards. Two or three lines. It is the same instinct as a version history, applied to editorial judgement rather than code.
US registration requires you to disclose more than de minimis AI content, and EU rules require labelling in defined circumstances from August 2026. Both are far easier to answer while the work is being made. Add one line to your creative brief template asking whether AI generated material will appear in the deliverable and who holds editorial responsibility for it. The answer helps identify which registration and transparency questions need separate treatment.
None of this is a reason to stop using these tools. The Copyright Office went out of its way to say that assistive use changes nothing, and it explicitly declined to recommend new legislation, concluding that existing law handles the question. The exposure is not in using AI. It is in assuming you hold rights without checking.
Return to one finished asset and compare it with the human source and edits. As with the cat’s unexpected hand, the key question is which expression the creator controlled and can evidence. The record is more useful than a count of how many prompts it took.
TL;DR , tap to expand the short version
- The US Copyright Office concluded in January 2025 that copyright does not extend to purely AI-generated content, and that prompts alone do not give a user enough control over the output to count as authorship.
- Three contributions do count, your copyrightable work perceptible in the output, creative selection and arrangement of AI generated material, and creative modification of it.
- The UK still protects computer generated works for 50 years under section 9(3) of the Copyright, Designs and Patents Act 1988. A government report published on 18 March 2026 recommended removing it.
- EU Article 50 separates provider marking from deployer labelling, with defined exceptions and transition arrangements. The maximum penalty figures are not automatic fines.
FAQ
Not the purely generated part. The US Copyright Office concluded in January 2025 that prompts alone do not give a user enough control over the output to constitute authorship. You can hold copyright in your expression that is perceptible in the output, in a creative selection and arrangement of generated material, and in creative modifications you make to it.
No. The Office addressed this directly and concluded that revising and resubmitting prompts generates more outputs to choose from without increasing your control over any of them. It also noted that copyright protects original authorship rather than effort, following the Supreme Court’s reasoning in Feist.
No. The Office was explicit that assistive uses do not affect copyright in the output, listing examples including colour correction, de-blurring, song ideation and generating a preliminary outline. Incorporating AI generated elements into a larger human authored work preserves protection for its original human expression, without necessarily protecting the generated components.
Not currently. Section 9(3) of the Copyright, Designs and Patents Act 1988 protects computer generated works with no human author for 50 years, assigning authorship to whoever made the arrangements necessary for creation. No reported UK case has applied it to generative AI, and a government report published on 18 March 2026 recommended removing the provision.
Article 50 transparency duties apply from 2 August 2026, with a marking grace period until December 2026 for systems placed on the market before that date. It distinguishes marking by providers from labelling of deepfakes and AI generated text on topics of public interest by deployers. The text labelling duty does not apply where content has undergone human review or editorial control and a person holds editorial responsibility for publication.
Spotted an error? Contact the desk. Our corrections policy explains how we handle corrections and updates.