In January 2025, the Getty Museum in Los Angeles announced the acquisition of its first artificial intelligence (AI)-generated photograph. The work, Cristian en el Amor de Calle by Matías Sauter Morera, was integrated into the exhibition The Queer Lens. Paul Martineau, the museum's curator, dismissed ethical objections: the artist mobilizes "hundreds of commands and choices that take months." At the same time, a few thousand kilometers away, Getty Images — a legally distinct entity from the museum but linked by name and family lineage — sued Stability AI in British and American courts for massive copyright infringement.
The same name. Two irreconcilable positions. This paradox encapsulates the legal chaos of an industry estimated at over $100 billion, operating without a compass.
Legal fronts multiply
The litigation meter is soaring. Nearly 60 lawsuits target generative AI companies. Five major trials are simultaneously underway in American and British jurisdictions. The pace of new complaints is not slowing down. It's accelerating.
The Getty Images vs. Stability AI case saw its first resolution in the UK on November 4, 2025. The judge ruled that the weights of an AI model — these numerical parameters that encode the knowledge acquired during training — do not constitute "copies" under British copyright law. Getty largely lost. A shockwave through the industry: if model weights are not copies, training on protected works might escape copyright enforcement. Getty Images refiled a complaint in California in August 2025, seeking a more favorable venue.
The Andersen vs. Stability AI case, filed in January 2023 as a class action, will reach the trial stage on September 8, 2026. Artists accuse Stability AI, Midjourney, and DeviantArt of having trained their models on protected works without consent or compensation. This now imminent date could produce the first substantive decision on the legality of training generative models in the United States.
Entertainment giants have opened their own front. Disney and Universal sued Midjourney on June 11, 2025, demanding $150,000 per work. The financial stakes, considering the volume of potentially affected works, run into billions. This is no longer a guerrilla war waged by independent artists. The studios have the resources to fight these battles to their conclusion.
When a freelance illustrator attacks Stability AI, the company can play for time. Against Disney's legal teams, that strategy no longer holds.
The Thomson Reuters precedent: end of innocence
A decision rendered in February 2025 in the Thomson Reuters vs. Ross Intelligence case drew a red line that the industry had not anticipated. The court rejected the defense of fair use — the mechanism in American law that allows, under certain conditions, the use of protected works without authorization — for training an AI system on proprietary content.
Until now, generative AI companies based their defense on four pillars: transformative nature of use, absence of direct reproduction, public benefit, and absence of commercial harm. Thomson Reuters vs. Ross shattered the first one. The judge ruled that using Westlaw content to train a competitor was not transformative. The signal is clear: fair use is not a universal pass for training AI models.
For the New York Times, which is suing OpenAI, this jurisprudence is valuable ammunition. In January 2026, a judge ordered the production of 20 million ChatGPT logs as part of the proceedings. The scale of this injunction demonstrates the seriousness with which American justice treats these disputes. Every conversation stored by OpenAI becomes potential evidence, likely to demonstrate that the model reproduces protected content. The stakes go beyond the Times's case: if the logs reveal systematic reproduction of protected articles, all press publishers would have a massive precedent to claim compensation.
Who is the author? An unanswered question
The Washington Circuit Court of Appeals ruled for the first time on March 18, 2025, in the case of Thaler v. Perlmutter. Stephen Thaler had filed a copyright application for an image entirely generated by his AI system, Creativity Machine. The rejection was confirmed: "Human authorship is required for copyright protection." The wording leaves no ambiguity. An AI alone cannot hold copyright.
The US Copyright Office hammered the point home in the second part of its report, published on January 29, 2025: "Prompts alone are insufficient to establish authorship." Formulating a textual instruction, however elaborate, does not make you the author of the produced image. The legal chasm is dizzying. If no one holds copyright on an AI-generated image, that image falls into the public domain upon its creation. Anyone can copy it, modify it, or commercialize it. For companies investing in the creation of AI-generated content — advertising campaigns, editorial illustrations, video game assets — the lack of protection represents a major commercial risk.
A competitor can freely use an AI-generated image without legal consequence.
The case of Zarya of the Dawn, decided in February 2023 by the US Copyright Office, outlined a middle ground. Kris Kashtanova obtained partial copyright on her comic book: human elements — selection, arrangement, text — were protected, while images generated by Midjourney were not. A "granular copyright," distinguishing layer by layer what constitutes human creation.
Jason Allen, creator of Théâtre d'Opéra Spatial — the Midjourney-generated work that won an art award in Colorado in 2022 — was denied copyright and is challenging the decision in court. His case embodies the distress of creators who invest dozens of hours in formulating, iterating, and curating generated images, only to be denied any legal paternity.
This debate has a precedent. In the mid-19th century, Nadar and his contemporaries fought in French courts to have photography recognized as an act of creation, not a mechanical product. The detractors' argument was identical: the machine does the work; the operator is merely an intermediary. It took decades for the law to adapt. History is stuttering, with one significant difference: generative AI does not have the luxury of waiting fifty years.
Three continents, three doctrines
International legal fragmentation adds a layer of complexity that global companies struggle to manage.
Japan has chosen maximum permissiveness. Article 30-4 of its copyright law, amended in 2018 and effective in 2019, authorizes the use of protected works for computer analysis — including the training of AI models — without notable restriction. Tokyo has made a strategic bet: attracting investment by offering the most welcoming legal environment in the world. AI companies find a haven there. Japanese creators, however, find themselves exposed to a mass scraping of their works without protection.
The European Union has opted for the regulatory path. The AI Act imposes, since August 2, 2025, transparency obligations on providers of general-purpose AI (GPAI) models. These companies must document the data used for training and comply with European copyright law. Effective enforcement is scheduled for August 2, 2026. The penalty for GPAI model providers — 3% of global annual turnover — gives a measure of the stakes: for a company valued at 100 billion euros, the fine could reach three billion. The approach attempts to reconcile innovation and protection, but its complexity raises fears of a competitive slowdown compared to Japan and the United States.
The United States leaves it to the courts to decide. Without specific federal legislation, the fair use doctrine remains the primary battlefield. But this mechanism, designed for individual cases, was never intended to arbitrate an industrial transformation of this magnitude. Each decision adds a piece to the puzzle without the overall picture becoming clear. The rejection of fair use in Thomson Reuters v. Ross contradicts the optimism of AI companies. The British decision on model weights in Getty v. Stability AI does not bind American judges.
This divergence creates a phenomenon of legal "forum shopping." AI companies seek the most favorable jurisdictions. Rights holders choose the most protective courts. Litigation is globalizing, but the law remains local. A company that trains its model in Japan, commercializes it from the United States, and serves European customers finds itself subject to three incompatible regimes. The absence of an international framework — no "AI Berne Convention" on the horizon — guarantees that this fragmentation will persist at least until the end of the decade.
Beyond art: the domino effect
Reducing this battle to the world of art alone would be a mistake. The precedents being set today in artistic disputes will define the rules applicable to all content used to train AI models. Journalistic texts, legal databases, computer code, musical compositions, medical data — everything is concerned. A judge who decides whether Midjourney has the right to ingest millions of illustrations indirectly sets the conditions under which a language model can absorb news articles or pharmaceutical patents. The industry relies on an unvalidated legal premise: that training data is available at no cost.
The $1.5 billion settlement agreed to by Anthropic in the Bartz v. Anthropic case (September 2025), concerning the use of pirated books for training its models, illustrates the magnitude of the amounts at stake. Nicole Sales Giles, director of digital art at Christie's, defends a more nuanced interpretation: "AI learns everything to create something new. That's influence. Not theft." This position rests on an analogy with human learning: a painter who studies the masters steals nothing from them.
Karla Ortiz, artist and leading figure in creators' resistance, responds directly: "Christie's simply doesn't care about artists." The opposition between these two visions is not merely philosophical. It reflects a conflict of economic interests between institutions that monetize art — auction houses, galleries, platforms — and those who produce it. Artists whose works feed the models see their market contract while the value they helped create enriches technology companies.
The Getty paradox, mentioned at the outset, embodies this tension. That the museum celebrates AI-assisted creation while the photo agency sues it is not an inconsistency: it is an accurate reflection of a world where the same actors are simultaneously beneficiaries and victims of disruption. The museum values the creative act of the artist who uses AI as a tool. The agency defends its catalog of millions of images scraped without permission. Both positions are defensible. Neither solves the problem.
The September 2026 wall
Five dates will structure the generative AI legal year. The Andersen v. Stability AI trial on September 8, 2026, will be the first real test of the legality of training on protected works in the United States. The entry into force of the enforcement of the European AI Act on August 2, 2026, will mark the beginning of a regime of real sanctions. The Disney and Universal v. Midjourney case will escalate. The New York Times v. OpenAI case, with its 20 million logs to analyze, may deliver its first conclusions on the reproduction of protected content. Jason Allen's appeal regarding Théâtre d'Opéra Spatial will continue to test the limits of copyright for hybrid works.
These decisions will not simply be technical. They will shape the distribution of value in an industry that already generates tens of billions of dollars in annual revenue. If American courts follow the Thomson Reuters logic and reject fair use for training, AI companies will have to negotiate massive licenses — or stop using protected content. If the British logic of "weights are not copies" prevails, creators will lose all legal leverage.
The most probable outcome lies between these two extremes. A compulsory licensing regime, comparable to that governing streaming music, could emerge. Creators would receive modest but systematic remuneration. AI companies would gain the legal certainty they demand. The details — remuneration rates, calculation mechanisms, arbitration bodies — remain to be invented. Anthropic's $1.5 billion settlement suggests that some companies prefer to pay rather than risk an unfavorable verdict. This transactional logic could set standards even before judges rule.
As with photography in Nadar's time, the law will eventually catch up with technology. The question is not whether a legal framework will emerge, but who will write it: judges, legislators, or the companies themselves, through billion-dollar transactional settlements. The answer lies in the next twelve months. It will concern anyone who produces, owns, or uses content — which is everyone.
This article is part of the "Artificial Intelligence and Creative Industries" dossier. See also: the analysis of the AI art market (article 1) and the investigation into the social divide caused by creative automation (article 4).
