AI and Copyright: Who Really Owns Works Created by Machines?
- Stéphane Guy
- 4 days ago
- 11 min read
The question has been circling since Midjourney generated images first flooded the internet in 2022. Three years on, the legal answer is beginning to take shape, and it's more uncomfortable than anyone anticipated: for creators, for AI companies, and for the everyday users of these tools.
An image generated in thirty seconds from a text prompt. A musical composition produced without human intervention. A novel written by a language model. These works exist, circulate, and sometimes sell. But their legal status remains, across most jurisdictions, a grey zone that neither legislators nor courts have fully illuminated.
In most countries, a work produced entirely by an AI without meaningful human creative input cannot, at present, be protected by copyright. Creation remains a human prerogative. But this apparently clear principle fractures into nuances, exceptions, and blind spots that make the question far more complex than it appears.
In Short
Neither in the US nor in the EU can an AI be recognized as an author: copyright law demands genuine human creative contribution. Fully autonomous AI works fall into the public domain.
"AI-assisted" works occupy a legal grey zone: if a human exercises real creative control over the final result (selection, iteration, post-editing), some protection may be available, but it is rarely total.
The Zarya of the Dawn case is an instructive illustration of the legal and artistic ambiguity surrounding AI-assisted creation.
Training data constitutes the second front of the conflict: can it be used without the consent of original creators? The Getty Images v. Stability AI ruling delivered partial and nuanced answers.
France is moving: a Senate bill passed in April 2026 establishes a presumption of use of cultural content by AI providers, reversing the burden of proof in favour of creators.
AI Creates. But Does It Really Create?
Start with the problem in precise terms. Copyright law, as codified in the United States under the Copyright Act of 1976, rests on a foundational principle: works of authorship. The Act protects "original works of authorship fixed in any tangible medium of expression."*
But here is the critical complication: a machine is not a legal person. It has no freedom of choice in any philosophically meaningful sense. It applies statistical probabilities to training data. What ChatGPT, Midjourney, or Suno produces is not, legally speaking, the expression of an original work of authorship. It is the output of a computation.
The logical consequence is harsh for many users: if you ask an AI to produce an image without contributing your own creative input, what you receive cannot be protected. Nobody owns that file. It falls, in theory, into the public domain. But is that really the case in practice?
United States: Thaler Loses, AI Remains Rightless
The case that most precisely defined the current legal landscape played out across the Atlantic, and it took years to resolve. Stephen Thaler, a computer scientist, developed an AI system he named the "Creativity Machine." This system autonomously produced a visual work titled A Recent Entrance to Paradise. Thaler filed a copyright registration application designating the machine as author, and himself as owner.
The Copyright Office denied it. The courts confirmed. On March 18, 2025, the U.S. Court of Appeals for the D.C. Circuit ruled definitively: “the Copyright Office had consistently interpreted the word “author” to mean a human prior to enactment of the 1976 Copyright Act.”*
Meanwhile, the U.S. Copyright Office published a landmark report in January 2025 addressing the copyrightability of works created with AI assistance. Its conclusion warrants close reading: prompts alone (the text instructions given to a generative AI) are insufficient to establish a copyright claim. The Office concluded that “prompts as unprotectible ideas or instructions”.*
In other words: formulating a request is not the same as creating. The idea remains in the public domain. The machine executes. The result belongs to no one.
One nuance worth noting: the report specifies that “AI generated content may be eligible for copyright “only if the human input in [the] creative process was significant.”” Having a good idea is not enough; you must realise it through your own means.*
The Grey Zone of "AI-Assisted" Works
None of this means that AI use automatically forecloses copyright protection. Legal scholars have spent the past two years hammering a key distinction: that between autonomous AI works and AI-assisted works.
Consider a photographer who uses an AI tool to retouch an image, sharpen it, or generate a background. If their creative contribution remains determinative, if they choose, select, and direct, the final result may be protected. American jurisprudence through the Copyright Office has begun drawing this boundary, recognizing partial rights over human-authored elements embedded in hybrid works.
The case of Kris Kashtanova illustrates this ambiguity with precision. Kashtanova created a graphic novel titled Zarya of the Dawn, using Midjourney for the visual elements while personally authoring the narrative, developing the characters, and structuring the story. It was, in other words, an AI-assisted work, one in which the technology served to realise a pre-existing, well-structured creative vision. Kashtanova initially obtained copyright registration, but the U.S. Copyright Office revoked it upon learning the images had been generated by Midjourney, a fact not disclosed in the original application.
However, the Office confirmed that the selection, coordination, and arrangement of the Work's written and visual elements will remain protected, as these were Kashtanova's own creative choices and not the AI's : “As a result, the new registration for Kashtanova’s work explicitly excluded “artwork generated by artificial intelligence.””*
This case is a textbook illustration of the grey zone surrounding AI-assisted creation. No clear, comprehensive legal framework governs this type of work, and one may never exist, given the sheer diversity of use cases, tools, and levels of human involvement. The rapid pace of AI capability improvements (and the increasing difficulty of detection, even as disclosure obligations multiply) makes any rigid legislative framework structurally fragile.
The Other Battle: Who Has the Right to Feed on Existing Works?
The ownership question over AI-generated outputs obscures a second, arguably more immediately concrete conflict for millions of artists and writers: the question of training data.
To produce photorealistic images, convincing prose, or coherent melodies, large AI models have ingested billions of human works, without asking permission, without paying licensing fees, and often without even notifying the original authors. This is where today's most intense legal battles are unfolding, and the hidden dangers of generative AI extend far beyond the image sector: books, journalism, photography, and music all face identical stakes. AI's impact on the music industry is only one chapter in a much broader story.
Getty Images v. Stability AI: A Muted Victory
In 2023, Getty Images sued Stability AI (creators of the Stable Diffusion model). The central claim: the unauthorized use of millions of Getty photographs (reportedly around 12 million works) scraped from Getty's databases to train the model, resulting in outputs that sometimes reproduced Getty watermarks and trade marks.
On November 4, 2025, the UK High Court handed down a ruling that was both anticipated and, in its details, surprising. It largely rejected Getty's claims. The court's central finding: Stable Diffusion's model weights do not store source images in any recognizable form, and the secondary copyright infringement claim failed on the basis that the model itself does not constitute an "infringing copy" of Getty's works under UK law.*
*Latham & Watkins, Getty Images v. Stability AI: English High Court Rejects Secondary Copyright Claim
However, the Court did find limited liability on the trademark front, noting that some versions of Stable Diffusion had produced outputs containing reproductions of Getty's iStock watermarks, constituting a narrow trade mark infringement.*
The verdict was perceived as a win for AI developers and a cold shower for copyright advocates. But it does not close every door: the Court expressly preserved Getty's primary copyright infringement claims for potential future determination, and parallel proceedings continue in the United States under a different legal framework.
What the case reveals above all is the technical difficulty of proving that a specific copyrighted work was used in a model's training and that the output constitutes a reproduction of it. Mass data scraping creates a diffuse harm, hard to demonstrate, near-impossible to quantify. It is precisely the kind of hidden danger of generative AI that legal practitioners and creators are beginning to name, without yet having the tools to measure it.
France Takes a Position (Cautiously, But It's Moving)
In France, the legal framework remains incomplete, but significant signals have emerged. In July 2025, the French Senate published a report titled "Création et IA : de la prédation au partage de la valeur" ("Creation and AI: From Predation to Value Sharing"), led by a commission under Senator Laurent Lafon.
Its graduated response recommendations, applicable after a government-led negotiation phase between AI providers and rights-holders, outlined three escalating steps:
Await the outcome of negotiations launched by the Ministries of Culture and Economy between AI providers and rights-holders.
In the event of failure, table a Senate bill implementing a presumption of use of cultural content by AI providers.
If that also fails, introduce a levy on the French revenues of AI providers and deployers, to compensate the cultural sector.*
France's position remains ambivalent, visible in every study and inquiry launched on the subject. Between the imperative of competitiveness (not penalizing European tech companies against American and Chinese giants) and the defense of creators' rights, the government is walking a tightrope.
The AI Act Solves (Almost) Nothing Here
Some observers had hoped that the EU's Artificial Intelligence Act, which entered into force on August 1, 2024, would clarify the intellectual property landscape. It did not. The AI Act focuses on AI system safety and fundamental rights protection. Under Article 53(1)(d), it requires “Providers create and publicly release a "sufficiently detailed public summary of the content used for the training of the model."”* But it does not adjudicate intellectual property ownership.*
The directive specifically dedicated to civil liability for AI, which could have opened the door to meaningful progress, was abandoned by the European Commission in February 2025, citing "no foreseeable agreement" among member states. A decision that speaks volumes about the political difficulty of legislating on a subject where economic and cultural interests are in direct contradiction.*
This European legal vacuum is precisely what the risks of AI and automation for our society analysis has long flagged: a fragmentary body of law, in permanent construction, struggling to track a technology whose capabilities evolve faster than legislative cycles.
Who Is Actually Coming Out Ahead?
In this legal landscape, who has the strongest position?
AI companies, for now. They operate in grey zones that tilt in their favour. Training on web-scraped data remains largely unsanctioned across most jurisdictions. Model weights do not store source works in recognizable form, the argument that persuaded the UK High Court. And the burden of proof still falls on creators.
Users occupy a more uncertain position. What you create with an AI may not fully belong to you, or may not belong to you at all, if your creative contribution is judged insufficient. The terms of service of generative AI platforms are often vague on this point.
Some are more explicit than others. OpenAI states in its Terms of Use: “As between you and OpenAI, and to the extent permitted by applicable law, you (a) retain your ownership rights in Input and (b) own the Output. We hereby assign to you all our right, title, and interest, if any, in and to Output.”*
In other words: you own what you put in and what comes out, but "to the extent permitted by applicable law," a qualifier that can mean everything or nothing, depending on the jurisdiction and the specific output.
Adobe takes a similar position for its Creative Cloud suite: “As between you and Adobe, you (as a Business User or a Personal User, as applicable) retain all rights and ownership of your Content. We do not claim any ownership rights to your Content.”*
Original creators, artists, photographers, writers, musicians, stand to lose the most in the short term, and to gain the most if the reforms currently underway produce results.
The Philosophical Question Everyone Avoids
There is a dimension that law, in its instrumental logic, struggles to accommodate: the question of the intrinsic value of human creation.
If an AI produces a symphony in four minutes that no one can distinguish from a human work, what is the value of the human symphony? If any internet user can generate in thirty seconds an image at the level of a professional illustrator, what does that say about the market for illustration? This is not simply a copyright question. It is a question about the economics of creative labour, and at its core, about what it means to be human.
Lawyers draw boundaries. But those boundaries shift with every model update. Tomorrow's AI will produce works that are even more indistinguishable, requiring even less human intervention. The moment when the current legal framework becomes obsolete may not be far away. If you want to understand what AI actually is and what drives these systems at a technical level, that context matters enormously for understanding why the law is struggling to keep up.
FAQ
Does an image I generate with Midjourney or DALL-E belong to me?
Not necessarily. In the United States, as confirmed by the Copyright Office and the D.C. Circuit Court in Thaler v. Perlmutter, if your creative contribution goes no further than the prompt itself, the image is not copyrightable. It falls into the public domain. If you exercised significant creative choices (iteration, post-editing, selection from outputs), partial protection may be available, but the legal line remains undefined.
Can a company file a patent on an invention conceived by an AI?
Not by designating the AI as inventor. In the United States, the Federal Circuit ruled as early as 2022 in Thaler v. Vidal that AI systems cannot be designated as inventors on a patent application. The European Patent Office adopts the same position. If a human can demonstrate a determinative creative contribution to the process, the patent can be granted to that person.
Does using copyrighted works to train an AI violate copyright law?
The question sits at the centre of multiple ongoing cases. The answer varies by jurisdiction and circumstance. In the UK, the High Court ruled in Getty Images v. Stability AI (November 2025) that Stable Diffusion's model weights do not constitute infringing copies of training works. In the United States, multiple cases remain pending. In Europe, Directive 2019/790 provides exceptions for text and data mining for research purposes, but their application to large commercial AI companies is vigorously contested.
What does the EU AI Act say about copyright?
Very little, precisely. Under Article 53(1)(d), the AI Act requires general-purpose AI model providers to publish a sufficiently detailed summary of content used for training. But it does not resolve questions of intellectual property ownership or creator remuneration. It is a transparency obligation, not a rights guarantee.
Will France's new law change anything for creators?
The bill unanimously adopted by the French Senate in April 2026 establishes a presumption of use of cultural content by AI providers. In practice: if an AI model could have used protected works, which is almost universally the case, the provider must prove it did not, rather than the rightsholder having to prove infringement. This reversal of the burden of proof significantly strengthens creators' litigation position, even if its practical enforcement remains to be built out. The bill now moves to the National Assembly.
Can you protect a work co-created with an AI?
In principle, yes, if the human contribution is sufficiently determinative. The U.S. Copyright Office recognises the copyrightability of human-authored elements in hybrid works, provided the author can clearly identify them. But the precise boundary between "AI tool use" and "co-authorship" has not yet been defined with clarity by the courts.
What happens if an AI unintentionally reproduces an existing work?
This is one of the most acute current risks. A model can generate an image, a melody, or a text that closely resembles a protected work without the user being aware of it. Liability may then shift toward the model provider, as illustrated by the Ultraman cases in China, where courts found AI platform operators contributorily liable for failing to implement adequate preventive filters.*
Comments