Generative AI can produce images, text, and music on demand. It can imitate a style closely enough to unsettle the artist who originated it. It can be trained on billions of copyrighted works scraped from the internet. And it has thrown copyright law — a body of rules designed for a world of human authorship and discrete copies — into deep uncertainty. Courts, regulators, and legislators are now grappling with questions that have no easy answers.
Can AI Be an Author?
The most fundamental question is whether a work created by AI can be copyrighted, and if so, by whom. The U.S. Copyright Office has taken a consistent position: copyright protection requires human authorship. In its March 2023 guidance, it stated that works generated entirely by a machine without human creative input are not copyrightable.
This position was tested in court. In Thaler v. Perlmutter, computer scientist Stephen Thaler argued that his AI system, “DABUS,” should be recognised as the author of an image it generated. The U.S. District Court for the District of Columbia ruled against him in 2023, holding that copyright law requires human authorship. In 2025, the D.C. Circuit Court of Appeals affirmed that ruling, cementing the human-authorship requirement in U.S. law. Similar conclusions have been reached in other jurisdictions.
The practical implication is that purely AI-generated content enters the public domain — it cannot be copyrighted, which means anyone can use it. This has significant consequences for the AI companies that hoped to monetise exclusive content and for users who assumed their generated images were protected.
Partial Human Authorship
The Copyright Office has been more nuanced about works that combine human and AI contributions. In its 2023 registration decision for Kris Kashtanova’s graphic novel Zarya of the Dawn, it granted copyright in the human-authored text and arrangement but denied protection for the individual AI-generated images, because the “expressive elements” were generated by the Midjourney system rather than by a human.
The Office has indicated that copyright can protect a human’s creative selection, arrangement, and modification of AI-generated material — the “sweat of the brow” applied by a human — but not the raw output. Defining exactly where that line falls is proving difficult and is likely to be litigated case by case.
The Training Data Question
The bigger legal battle concerns the training data itself. Did AI companies commit copyright infringement by training models on copyrighted works without permission or compensation? Several major lawsuits are testing this.
Getty Images sued Stability AI in both the U.S. and the UK, alleging that Stability scraped millions of Getty’s images (including watermarked ones) to train Stable Diffusion, and that the model can reproduce Getty’s distinctive watermark. A group of visual artists filed a class action against Stability AI, Midjourney, and DeviantArt. Authors, including a group led by the Authors Guild, sued OpenAI over the use of their books in training ChatGPT. Music publishers and record labels have sued AI music generators Suno and Udio.
The core legal questions are whether training constitutes “fair use” under U.S. law and whether training on copyrighted works violates the reproduction and derivative-work rights of rights-holders. Fair use analysis considers the purpose and character of the use, the nature of the work, the amount used, and the effect on the market. AI companies argue training is transformative, produces new works, and does not substitute for the originals. Plaintiffs argue training is commercial, copies the entire work, and harms existing and potential markets for licensing.
Emerging Judicial Signals
Early rulings have gone both ways. In 2023, a federal judge in the Andersen v. Stability AI case allowed most of the artists’ claims to proceed, finding that allegations of copyright infringement were plausible, though he rejected some. In the Authors Guild case, a judge allowed the core copyright claims to proceed, narrowing but not dismissing them.
Courts have been cautious about ruling on fair use at the preliminary stage, indicating that full trials may be needed. That means definitive answers may take years.
The Copyright Office’s Reports
The U.S. Copyright Office, tasked by Congress with studying the issue, has issued reports on digital replicas, copyrightability, and the training question. Its reports have generally taken a rights-holder-friendly tone on training, suggesting that using copyrighted works to train models may not automatically qualify as fair use and that market harm must be assessed in the context of a functioning licensing market.
Policy and Legislative Responses
Legislatures are moving, unevenly. The EU AI Act includes a transparency provision requiring providers of general-purpose models to disclose a “sufficiently detailed summary” of training data and to respect machine-readable opt-outs for text and data mining. Several U.S. states have proposed laws. Canada’s AIDA, which would have addressed some AI issues, died with prorogation in 2025, leaving a gap.
Meanwhile, a de facto licensing market is emerging: news organisations, authors, and image libraries have struck deals with AI companies to license content for training. These deals suggest the market is finding a value, even as the law remains unsettled.
What This Means for Creators
For artists, writers, and musicians, the practical landscape is contradictory. AI can generate content that competes with their work, trained on their work, often without compensation. Yet the law currently denies copyright to purely AI-generated output, which limits how AI companies can monopolise the content they generate and leaves room for human authorship to retain value.
The most likely near-term outcome is a mix of: continued litigation, a growing licensing market, transparency requirements, and evolving norms about disclosure of AI use. A sweeping legislative rewrite of copyright for the AI era is unlikely in the near term, especially in the ideologically divided United States.
The Fair Use Fight in Detail
The fair-use defence turns on four statutory factors, and each is contested. On purpose and character, AI companies argue training is “transformative” because it uses works to extract statistical patterns for a new purpose, not to reproduce them. Plaintiffs counter that commercial training that produces competing outputs is not transformative in the sense courts have recognised. On the nature of the work, the fact that training uses published, expressive works cuts against fair use. On amount used, copying entire works wholesale — as scraping does — weighs against the defence, though companies note the individual work is a tiny fraction of the training set. On market effect — often decisive — the fight is over whether AI outputs substitute for the originals and whether training displaces a potential licensing market. The Copyright Office has leaned toward finding market harm where licensing markets exist or could exist.
Style, Likeness, and the Right of Publicity
Copyright is not the only legal tool. Artists whose styles are imitated may have weak copyright claims — style itself is not protectable — but they may have rights of publicity, trademark, or unfair-competition claims, particularly where AI generates content that falsely implies endorsement. Voice actors and musicians have pursued likeness claims against AI voice-cloning tools. The federal ELVIS Act in Tennessee and similar state laws target unauthorised voice and likeness replication. These adjacent rights are becoming as important as copyright in the fight over AI-generated content.
The Licensing Market Emerges
Even as litigation grinds on, a market is forming. OpenAI has signed deals with news organisations including the Associated Press, Axel Springer, and Le Monde, and with stock-image and content providers. Shutterstock and Getty have launched AI-training licences. Individual authors and artists can opt in to collective licensing schemes. These deals are imperfect — critics note they often compensate rights-holders modestly and may legitimise past uncompensated use — but they demonstrate that courts are not the only mechanism for resolving the conflict. A functioning licensing market could, in time, make the fair-use question less existential.
Transparency and Provenance
A parallel strategy is technical: content provenance. Standards like C2PA (Content Credentials) embed cryptographically signed metadata that records how an image or video was created, including whether AI was involved. Platforms and camera makers are adopting them. The goal is to let viewers distinguish authentic from synthetic media and to let rights-holders track use. Provenance will not resolve copyright disputes, but it can make the provenance of training data and generated output more legible, which matters for both enforcement and public trust.
What Creators Should Do
For individual creators, the practical takeaways are mixed. Registering human-authored works remains valuable, and documenting the human creative contribution strengthens claims where AI tools were used. Opting out where platforms offer it, joining collective licensing efforts, and monitoring for unauthorised use are prudent. But the deeper truth is that the legal landscape is unsettled, and no strategy guarantees protection. The most robust position is to create work whose value lies in what is hardest to replicate — voice, judgement, lived experience, and the human relationships that no model possesses.
Conclusion
The convergence of generative AI and copyright has produced one of the most consequential legal questions of the decade. The direction of travel — human authorship required for protection, training data under scrutiny, licensing markets emerging — is becoming clearer, but the details remain contested. The outcomes will determine who owns the output of the coming creative revolution, and whether human creators share in its value.
