Suppose you are making a poster.
You decide the subject, the mood, the message, and the audience. You ask an AI system for twenty visual directions. Nineteen are wrong. The twentieth contains one useful idea, so you keep the composition but change the colors, redraw the face, replace the background, rewrite the text, move every element, and discard half of what the system produced.
By the end, the poster does not look like any single AI output you saw along the way.
You publish it.
Then someone asks a question that sounds simple:
Who made this?
You did not create every pixel directly. The model did not choose the purpose of the work, decide which version mattered, or take responsibility for the final result. The software company built the system. Millions of earlier works may have shaped what the model could generate.
One file now contains several different kinds of contribution.
So who owns it?
When creation becomes a chain of contributions, ownership stops being a single question.
A prompt can contain an idea without controlling the expression
This is the part that frustrates many people who work carefully with generative systems.
A prompt can be detailed. It can specify a scene, a palette, a camera angle, a rhythm, a genre, a character, a sequence of events, or a page of constraints.
Why should that not count as authorship?
The U.S. Copyright Office’s answer is not that prompts are meaningless. It is that, with current generative systems, prompts generally do not determine the resulting expressive elements closely enough.
You can describe exactly what you want and still receive something you did not predict. Change one word and the system may alter many unrelated details. Repeat the same prompt and you may receive a materially different result.
The gap between instruction and output matters.
If I tell a human illustrator, “Draw a child standing beside a red bicycle under a broken streetlight,” I supplied ideas and constraints. The illustrator may still decide the child’s face, posture, line weight, framing, shadows, texture, and thousands of other expressive details.
Giving the instruction did not make me the sole author of the illustration.
A generative model complicates the analogy because there is no second human author on the other side. But the underlying question remains useful:
Who actually determined the expressive form that appears in the final work?
The answer can change as tools change. A future system might allow a person to control expressive elements with enough precision that the legal analysis changes. The Copyright Office explicitly leaves room for that possibility.
So the boundary is not “AI touched it, therefore no copyright.”
The boundary is about human control over expression.
U.S. Copyright Office (2025), Part 2: prompts and control over expressive elements
Selection can itself become creation
Now return to the poster.
Suppose the model produces one hundred images. You choose six. You crop them, combine parts, alter the lighting, paint over details, add typography, and arrange the final composition yourself.
Even if some underlying AI-generated fragments are not copyrightable in the United States, your selection and arrangement may be.
The Copyright Office reached a related result in Zarya of the Dawn, a graphic work that combined human-written text with Midjourney-generated images. The Office limited the registration: it recognized the human-authored text and the human selection, coordination, and arrangement of the material, while excluding the AI-generated images themselves.
That sounds technical, but it carries a larger idea.
A work does not have to be legally all-or-nothing.
One layer can be protected while another is not.
The same logic already exists in older creative forms. A collage can contain public-domain photographs and still receive protection for an original arrangement. An edited anthology can contain preexisting works while adding protectable selection and organization. A film can contain music licensed from someone else without transferring authorship of the music to the director.
Human–AI work makes layered authorship more common, not conceptually new.
The difficult part is documenting where the human choices actually are.
U.S. Copyright Office (2023), Zarya of the Dawn registration decision
Follow the Question: Creative Agency
The machine can contribute without becoming an owner
There is another assumption hidden inside the word collaboration.
If two things contribute, perhaps both should own the result.
That is how we often think about human collaboration. Two writers draft a screenplay together. Two composers build a song. Two engineers jointly author code. Their contributions can create joint rights.
But legal ownership requires more than causal contribution.
A language model is not currently treated as a legal person capable of holding copyright, signing a contract, receiving royalties, assigning rights, or suing an infringer in its own name.
That does not mean its contribution is imaginary.
A system can generate a phrase you would not have thought of. It can suggest a visual structure that changes the entire work. It can introduce a solution that becomes central.
Causal importance and legal authorship are different categories.
We already accept that distinction elsewhere. A telescope can make a scientific discovery possible without becoming a scientist. A search engine can surface the source that changes an essay without becoming a co-author. Software can perform millions of calculations that shape a building without becoming the architect.
Generative AI feels different because the contribution resembles human expressive behavior.
That resemblance makes attribution psychologically difficult even when the legal rule remains human-centered.
Follow the Question: Production
The same artifact can contain several different rights
Consider a three-minute animated film made by one person with AI tools.
The story was written by the filmmaker. A model generated background concepts. The filmmaker drew the final characters. Another model generated temporary music. A human composer replaced most of it but retained one AI-generated texture. The filmmaker edited the footage, recorded the narration, and licensed a typeface.
Who owns the film?
There may be no single answer.
The screenplay can have one copyright. The drawings can have another. The recording can carry separate rights from the underlying music. The final editing can contain human authorship. Some AI-generated elements may have no copyright protection at all in one jurisdiction. Contract terms may govern access to the software or output even where copyright does not.
This is why asking “Who owns the AI work?” can be too coarse.
There may be no legally unified thing called the AI work.
There is a bundle of contributions, permissions, exclusions, licenses, and creative decisions assembled into one artifact.
For creators and businesses, that means provenance becomes valuable.
Which parts were written by a person? Which were generated? Which were modified? Which came from licensed material? Which contributor assigned rights? Which tool terms applied at the time?
The more commercially important the work becomes, the more expensive it can be to reconstruct those answers later.
Follow the Question: Contracts
A platform’s promise is not the same thing as copyright
Generative services often use contract language about outputs: who may use them, whether the provider claims rights in them, what restrictions apply, and what happens when multiple users receive similar material.
Those terms matter.
But a contract cannot create copyright in expression that copyright law does not protect.
If a platform says that, as between you and the platform, it assigns whatever rights it has in an output to you, that tells you something important about the relationship between you and the platform.
It does not necessarily tell you whether the output contains copyrightable authorship, whether someone else’s rights are implicated, or whether another jurisdiction reaches the same result.
This distinction is easy to miss because both questions are expressed with the same everyday sentence:
“Do I own this?”
Contract can answer who has promised what to whom.
Copyright answers a different question: what exclusive rights exist under law, and in which expressive elements.
A creator who relies on AI commercially needs both answers.
Follow the Question: Training & Provenance
Training data changes the question again
Suppose your final image is sufficiently shaped by your own creative decisions to qualify for copyright.
Does that settle whether the image was lawfully created?
No.
Authorship of an output and legality of the model’s training process are separate questions.
The U.S. Copyright Office’s 2025 Part 3 report on generative-AI training treats training as a copyright question of its own. It explains that training can involve acts within copyright owners’ exclusive rights and that fair use depends on the facts, including the purpose of the use, the source material, the nature of the resulting model, and market effects. The Office specifically distinguished potentially transformative training uses from situations involving pirated source libraries or outputs that substitute for protected works.
The European Union has taken a different regulatory route in part. Under the AI Act, providers of general-purpose AI models must maintain a policy to comply with Union copyright law and must publish a sufficiently detailed summary of training content. The obligations for new general-purpose models have applied since August 2, 2025, with enforcement available from August 2, 2026.
Those rules do not decide who owns every output.
They show why ownership cannot be isolated from provenance forever.
A person can have a strong claim to authorship in a final work while legitimate disputes remain about what material was used upstream to make the model capable of producing it.
U.S. Copyright Office (2025), Copyright and Artificial Intelligence, Part 3: Generative AI Training
European Commission AI Act Service Desk, Article 53 obligations for general-purpose AI models
European Commission, public training-content summary requirements for general-purpose AI models
Follow the Question: Comparative Law
Different countries can attach ownership to different places
There is no single global answer.
The United States currently rejects copyright in purely AI-generated material without sufficient human authorship.
The European Union likewise does not provide a special copyright right for works with no human author; its originality standard is tied to an author’s own intellectual creation and free and creative choices.
The United Kingdom is unusual.
Section 9(3) of the Copyright, Designs and Patents Act provides a category for certain computer-generated literary, dramatic, musical, and artistic works made without a human author. For those works, the law treats the author as the person who made the arrangements necessary for the work’s creation, and the protection lasts fifty years from creation.
But even that rule is unsettled in direction.
In its March 18, 2026 report on copyright and AI, the UK government said there is little evidence that this special computer-generated-work protection is actively used or materially supports creativity and innovation. It proposed that the provision should be removed in the absence of evidence of ongoing value, while continuing to monitor the issue.
So a work generated through the same process may encounter materially different legal frameworks depending on where protection is claimed.
That is not a small technicality for creators who publish globally.
The internet distributes one file everywhere. Copyright law still arrives by jurisdiction.
Follow the Question: Attribution
Credit is not the same thing as ownership
Law can tell us who may hold exclusive rights. It does not completely answer who deserves acknowledgment.
That is partly a social question.
A 2025 CHI study of 155 knowledge workers asked how people assign credit in human–AI co-creation. Participants did not treat every AI contribution equally. Their judgments changed with the type of contribution, how much the system contributed, and how much initiative it took. They also tended to assign AI less credit than a human making an equivalent contribution, while many still considered disclosure of AI involvement important.
Other research suggests that audiences react not just to the final artifact but to the process behind it. Experiments on AI-assisted visual art found that people often evaluated work differently when they learned AI had been involved, particularly when AI was used in implementation rather than only in ideation. Perceived authenticity helped explain some of that difference.
This creates a tension.
Legal authorship asks where protectable human expression exists.
Audience judgment may ask something closer to: how much of the difficult, intentional, identity-bearing work did the person actually do?
Those are not identical standards.
A creator can legally own a work and still face a credibility problem if the public believes the human contribution was overstated.
Conversely, someone can make a genuinely thoughtful human contribution to a process even when copyright protection for some resulting elements is thin or uncertain.
Credit lives partly in the story of how the work was made.
The useful question is not how much AI was used
People often want a percentage.
Was this twenty percent AI? Fifty percent? Ninety percent?
That number feels objective. It usually is not.
A model can generate ninety percent of the visible words while a human determines the entire structure, evidence standard, argument, and final editing. Another person can type ninety percent of the words manually while following an AI-generated outline they barely questioned.
Surface quantity does not map neatly onto creative control.
A better practical inquiry follows the decisions.
Who defined the purpose?
Who chose the constraints?
Who rejected alternatives?
Who introduced the details that make this work rather than any other plausible work?
Who changed direction when the result was wrong?
Who selected what survived into the final version?
Who can explain why the final choices were made?
Who accepts responsibility for the result?
This is not a statutory copyright test. Courts and copyright offices apply legal standards, not an editorial checklist.
But the questions help locate human agency.
They also reveal why two projects that both say “made with AI” can involve radically different creative processes.
Follow the Question: Creative Practice
Ownership may become a record of decisions
For most of creative history, the finished object carried much of the evidence of authorship.
A manuscript existed in drafts. A painting showed brushwork. A photographer had negatives. A software repository recorded commits.
Generative systems can compress creation into a sequence that is harder to see from the final artifact.
A finished image may conceal two hundred rejected generations and six hours of human editing. A polished essay may conceal the opposite: one prompt and almost no meaningful revision.
The outputs can look equally finished.
That makes process records more valuable.
Drafts, edit histories, source files, prompt iterations, masks, layers, code commits, notes, sketches, and version histories can show where judgment entered the work.
That documentation can matter for copyright registration, contractual disputes, internal company ownership, collaboration, and simple credibility.
The future of authorship may therefore become strangely procedural.
To show that you created something, you may increasingly need to show how decisions moved through the system.
Where the Fields Collide
Copyright law asks whether protected expression comes from a human author.
Human–computer interaction asks how much initiative, control, and creative direction move between person and system.
Contract law determines what users and platforms promise each other even when copyright protection is uncertain.
Economics asks what kinds of exclusive rights are actually needed to encourage valuable creation when machine-generated output can be produced at enormous scale.
Ethics asks who deserves credit, whose labor remains invisible, and whether disclosure changes the meaning audiences attach to a work.
AI governance adds provenance: what data trained the system, what rights were reserved, and what obligations attach upstream.
The fields collide because creation is no longer always a moment.
It is a pipeline.
Ideas enter from one place. Models transform them. Human judgment redirects them. Other people’s works may sit in the history of the system. Contracts surround the tool. Law protects some layers and not others.
Ownership becomes easier to understand when we stop looking for one magical point where the work was made and start tracing the decisions that gave it its final form.
The important boundary may not be human versus machine. It may be intention versus automation inside the same creative process.
What We Know — and What We Don't
We know that, in the United States, purely AI-generated expression without sufficient human authorship is not protected by copyright under the current approach of the Copyright Office and the D.C. Circuit.
We know that human-authored contributions can remain protected when AI is used as a tool, and that creative selection, arrangement, and modification can matter.
We know that the Supreme Court’s March 2026 denial of review in Thaler did not create a new nationwide opinion of its own; it left the lower-court ruling in place.
We know that jurisdictions differ. The UK still has a statutory computer-generated-work provision, even as the government has proposed removing it absent evidence of continuing value.
We know that training-data legality, output copyrightability, platform contracts, and social credit are separate questions.
What we do not have is a universal percentage of human contribution that automatically creates authorship.
We do not know how courts will apply existing rules to every new interface that gives users more granular control over generation.
We do not know whether lawmakers will eventually create new rights for machine-generated output, narrow existing rights, or leave the present human-authorship framework largely intact.
And we do not know whether audiences will settle on durable norms for disclosure and credit.
The technology is moving faster than the conventions around it.
That makes precision more useful than certainty.
Back to the Poster
Return to the poster on your screen.
The AI proposed possibilities. You chose one direction. You rejected others. You rewrote, redrew, rearranged, and decided when the work was finished.
The cleanest answer may not be “the human owns everything” or “the AI made it.”
The better answer is layered.
Some expressive elements may be yours because your creative decisions shaped them.
Some generated elements may not be protected at all under U.S. copyright law.
A platform contract may give you broad rights to use the output without changing that copyright analysis.
Other rights may belong to collaborators, licensors, or earlier creators whose protected material appears in the final work.
And the AI system itself does not become a copyright owner merely because it contributed something surprising.
So when someone asks, “Who made this?” the legally accurate answer and the socially satisfying answer may both require a story about the process.
Perhaps that is the real change.
Authorship used to sound like a name.
Increasingly, it may look like a map of decisions.
The Next Question
If authorship becomes difficult when one work contains several sources of agency, identity becomes even stranger when one person can be copied.
Suppose a digital system could reproduce your memories, habits, voice, preferences, and personality so precisely that the copy insisted it was you.
The original continues living. The copy does too.
Which one would be you?
Sources & Further Reading
- U.S. Copyright Office (2025), Copyright and Artificial Intelligence, Part 2: Copyrightability
- U.S. Court of Appeals for the D.C. Circuit (2025), Thaler v. Perlmutter
- U.S. Copyright Office testimony to Congress (May 12, 2026), noting the Supreme Court denial of review in Thaler
- U.S. Copyright Office (2023), Zarya of the Dawn registration decision
- U.S. Copyright Office (2025), Copyright and Artificial Intelligence, Part 3: Generative AI Training
- European Commission AI Act Service Desk, Article 53 obligations for general-purpose AI models
- European Commission, public training-content summary requirements for general-purpose AI models
- UK Government (March 18, 2026), Report on Copyright and Artificial Intelligence — computer-generated works
- He, Houde & Weisz (CHI 2025), Which Contributions Deserve Credit? Perceptions of Attribution in Human-AI Co-Creation
- Co-creating art with generative artificial intelligence (2024), Computers in Human Behavior: Artificial Humans
Beyond the Question is an interdisciplinary series by Arin Vale.
Read the editorial approachTHE NEXT QUESTION
If a perfect digital copy of you existed, which one would be you?
Coming next in Beyond the Question.
Explore the series