Artists Keep Winning the Case yet, Losing the War
Introduction : My dream of becoming an artist was hushed down when the people around me echoed that it’s not a safe career for earning enough to sustain. I still think of pursuing art but those words are now working out like prophecy, not because artists lack talent, but because the AI systems now learning from their work were never asked for permission.
Every few months, a court somewhere in the world is asked the same question, “Can an artificial intelligence system own the copyright to what it creates?” The ratio always comes back as “No”. In the very recent, India's Copyright Office affirmed the same in not granting DABUS the authorship title in 2026. The United States said so in Thaler v. Perlmutter. China's Beijing Internet Court said so in Li v. Liu. On paper, this looks like a settlement but in practice, it is settling the wrong dispute.
Even though it is settled that AI cannot be the holder of Copyright alone, we are missing the point of when courts start granting co authorship status under technicality and the other supposed human copyright holder starts monetising from it. With the way these neural networks work, they are about whether a human artist's work, a photograph, a painting, a distinctive style built over years can be absorbed, replicated, or trained upon without consent, credit, or compensation. Zhang Jingna's fight over her own photograph in Luxembourg, and the ongoing training-data lawsuits against Stability AI, Getty Images' opponents, and OpenAI, all sit on this messy battlefield.
A Recent Entrance to Paradise: The emerging path to peril?
Dr. Stephen Thaler, the creator of DABUS (Device for the Autonomous Bootstrapping of Unified Sentience) has tried to file the machine as the sole author of an artwork titled "A Recent Entrance to Paradise” across multiple jurisdictions, namely the United States, the United Kingdom, Australia, and, most recently, India, where the Copyright Office rejected the application on September 1st 2026. The reasoning across jurisdictions has been remarkably consistent as courts have refused non-human authorship before, most famously in Naruto v. Slater, where a macaque's self-taken photograph was denied copyright because animals cannot hold statutory rights. Authorship, in every one of these rulings, is a legal status reserved for persons, natural or juristic, and DABUS is neither. Except in the patent version of the DABUS saga in Australia, Justice Beach of the Federal Court initially ruled in Thaler's favour, allowing DABUS to be named as an inventor on a patent application. That ruling was reversed on appeal by the Full Federal Court, restoring the more familiar position, but for a short window Australia was the one country where an AI system had judicial recognition as a creator of intellectual property, even if only in the patent context rather than copyright.
United Kingdom's Copyright, Designs and Patents Act 1988 contains Section 9(3), grants authorship of a computer-generated work, one produced "in circumstances such that there is no human author," to "the person by whom the arrangements necessary for the creation of the work are undertaken." It is a workaround rather than a concession to machine authorship, since it still requires a human to be named, but it is the one route by which a fully autonomous AI output can receive protection in the UK without anyone needing to claim they personally created it in the ordinary sense.
None of these rulings stated the artwork lacked originality. India's order, in fact, appeared to accept that the work cleared the threshold of originality under Section 13 of the Copyright Act. The problem was never the art. It was the applicant.
This is worth sitting with for a moment, because Thaler's own account of how DABUS worked undercuts his own claim. He has acknowledged feeding the system his own original photographs as training material, which arguably makes him and not “DABUS” the more obvious candidate for authorship, in the same way a photographer who builds a machine to remix his own negatives would still be understood as the author of what comes out. Framing the output as DABUS's work rather than his own was, in that sense, a choice, not a necessity. It raises an uncomfortable question presently now that what if in these cases, the training material is not the developer's own photographs, but images scraped from across the internet, without the knowledge of the people who made them?
Zhang Jingna and the Painting in Luxembourg
Photographer Zhang Jingna,case did not involve AI at all but exposed the same underlying vulnerability which is why there is a need to have an answer to this question more than ever. A painter, Jeff Dieschburg, used one of Zhang's fashion photographs as visual reference for a commissioned painting, which went on to be exhibited and even win a prize. When Zhang objected, the case first went against her but the appellate court found that the painting did constitute an infringement of Zhang's copyright, regardless of the medium the appropriation took, and regardless of the fact that the photograph had been publicly available online.
This essentially clarified that “availability” is not the same as consent. A photograph being visible on the internet does not put it in the public domain, and a court had to say this explicitly, in 2024, about a single painter copying a single photograph. If that principle needs restating for one human artist referencing one image, it is hard to see how it survives contact with a generative model trained on hundreds of millions of images, scraped indiscriminately, with no realistic way for any individual photographer to even know their work was used, especially when the bar for proving originality is so low!
The US Supreme Court settled this decades ago in Feist Publications v. Rural Telephone Service, holding that copyright protection requires only a modicum of creativity, not novelty or artistic merit. Zhang Jingna's photograph, with its deliberate lighting, styling, and composition, should never have needed a second look to clear that bar. If the first Luxembourg court failed to appreciate that a single, well-staged fashion shot is original, it is not hard to imagine that they would not be able to consider originality and appropriation on a scale of millions of images that run through a model that nobody but the company can thoroughly access.
Same Blind Spot, Larger Scale: The Training-Data Lawsuits
Even when authorship is denied to a machine, a generative model can still be trained with an artist's entire body of work, learn the brush stroke of that artist or the lighting of that artist, and then be able to produce something that is close enough to the original work to be in competition with the original artist without the artist's name on it. The authorship question and the appropriation question are, legally speaking, entirely separate. Courts have now answered the first one clearly. They are still struggling with the second.
Andersen v. Stability AI is the clearest illustration in the United States. A group of visual artists filed a lawsuit against Stability AI, Midjourney, and DeviantArt, alleging that their copyrighted artwork was used without permission or compensation in training their diffusion models. Getty Images has pursued a parallel track against Stability AI in both the United States and the United Kingdom. The UK case gave a result that captures exactly how unresolved this area still is. The central argument, that the AI model itself was storing infringing copies, was rejected by the English High Court in November 2025, but the narrower argument of infringement due to the use of Getty's own watermarks was upheld. It was a case that did show that artists' work had clearly got into the model, but it did not result in a ruling on whether it was illegal or not.
Germany offered a more decisive answer, though only for text. In November 2025, the Munich Regional Court ruled that ChatGPT had memorised and reproduced the lyrics of nine well-known German songs, in the case brought by the music-rights body GEMA against OpenAI. That is one of the few rulings anywhere to find, on the facts, that a generative model did not just learn from copyrighted material but reproduced it. For visual art, no comparable finding has yet landed in a major jurisdiction which means artists like Zhang Jingna are in a more disadvantageous position than songwriters.
When Training Actually Lost: Thomson Reuters v. Ross Intelligence
Most of the training-data litigation discussed so far is still pending, which makes one decided case unusually important. In February 2025, a US federal court ruled against Ross Intelligence, a legal-research AI company, finding that its use of Thomson Reuters' Westlaw headnotes to train its models was not protected by fair use. This is the detail the Getty and Andersen cases have so far avoided. A court actually weighed the "transformative use" defence, the same defence generative-AI companies rely on, and rejected it on facts closer to home for anyone in the legal profession. It does not automatically decide how a court will rule on a diffusion model trained on photographs rather than headnotes, but it removes the assumption that training an AI on someone else's copyrighted work is presumptively transformative. If headnotes were not transformative enough, photographs, illustrations, and paintings may not be either.
ANI Media v. OpenAI
On 24 July 2026, Justice Amit Bansal of the Delhi High Court dismissed the interim injunction Asian News International sought against OpenAI, in a 135-page order that is the first substantive Indian judicial engagement with AI training and copyright. ANI's case rested on two claims: that ChatGPT was trained on its news content without a licence, and that it sometimes reproduced that content or invented quotes and falsely attributed them to ANI.
On training, ANI could not actually show its content sat in ChatGPT's training data - its examples postdated the models' training cut-offs. The court reached Section 52(1)(a) fair dealing only because OpenAI admitted to temporarily storing ANI's content during training; on that narrower, admitted fact, the court held the storage was, prima facie, fair dealing. On reproduction, ANI simply failed to show ChatGPT's outputs reproduced its articles, so that claim never reached the fair-dealing analysis at all. The hallucination claims that ChatGPT invented stories and attributed them to ANI, a reputational harm distinct from copying not decided either way. The court did, however, affirm it has territorial jurisdiction over OpenAI despite the company's offshore servers, which matters for every future Indian claim against a foreign AI developer.
Unlike with the easier question in DABUS, ANI V OpenAI asked the harder question I.e whether training an AI system on copyrighted material, without a licence, is something Indian copyright law will tolerate. The ruling favoured OpenAI, on a technicality, without resolving whether training on copyrighted content is actually lawful.
What China's Courts Reveal About the Missing Middle Ground
China's courts have, somewhat unexpectedly, done more to fill this gap than most jurisdictions. In Li v. Liu, the Beijing Internet Court held, just as the United States and India would later hold in the DABUS matter, that the AI tool itself could not be an author. But instead of stopping there, the court went on to find that the human user who iteratively prompted and refined the image had exercised enough creative judgment to be its author himself, and ruled in his favour when someone else lifted the image without credit. That is the missing middle ground DABUS never reached, because Thaler insisted on naming the machine as author instead of himself. Li v. Liu suggests that courts are willing to protect AI-assisted human creativity, they simply need the human to claim it.
For a working photographer, illustrator, or painter, that distinction is not academic. Nobody was ever going to sue on their behalf to stop a machine from being named an author, that was never the threat. The threat was always what happened to Zhang Jingna. Someone else, human or machine, absorbing a distinctive body of work and reproducing it commercially, while the original creator is left to litigate for years just to be recognised as the source. DABUS closes off one narrow avenue of harm but leaves the far larger one completely open.
Conclusion
The consensus against AI authorship is real, largely consistent, and, as far as it goes, correct - even the UK's Section 9(3) workaround and Australia's brief reversal only bend the rule rather than break it, since both still require a named human to stand behind the work. DABUS should not have been named an author, and almost no jurisdiction has seriously suggested otherwise. But treating that consensus as a solution to the broader anxiety around generative AI mistakes a procedural victory for a substantive one. Zhang Jingna had to fight, and initially lose, to establish that a human artist copying her photograph into a painting was still infringement , a proposition that Feist should have settled decades earlier. Getty Images fought Stability AI in two countries and, in the UK at least, walked away with a trademark ruling instead of an answer to the training-data question it actually cared about. Thomson Reuters v. Ross Intelligence shows that when a court does confront the fair-use question directly, the training defence does not automatically win. And with Andersen v. Stability AI headed to trial and ANI Media v. OpenAI now pending before the Delhi High Court, the next real test of this area will not be whether a machine can be an author. It will be whether the humans who trained it can keep using other people's work without asking. Until courts are willing to scrutinise that question with the same rigour they brought to the DABUS authorship question, artists will keep winning the argument that machines cannot own their work, while losing the argument that their work cannot be taken without their say.
Author: Adya Annapurna Sarangi in case of any queries please contact/write back to us via email to content@khuranaandkhurana.com or at Khurana & Khurana, Advocates and IP Attorney
References
Thaler v. Registrar of Copyrights, Copyright Office, Government of India (2026) — order rejecting AI authorship for "A Recent Entrance to Paradise."
Thaler v. Perlmutter, U.S. District Court for the District of Columbia (2023), affirmed on appeal.
Li v. Liu, Beijing Internet Court, (2023) Jing 0491 Min Chu No. 11279 (27 November 2023).
Zhang Jingna v. Dieschburg, District Court of Luxembourg (first instance, 2022) and Court of Appeal of Luxembourg (2024).
Andersen v. Stability AI Ltd., et al., U.S. District Court for the Northern District of California, No. 3:23-cv-00201 (ongoing; trial scheduled September 2026).
Getty Images (US), Inc. v. Stability AI, Inc., U.S. District Court for the District of Delaware (ongoing).
Getty Images v. Stability AI Ltd., England and Wales High Court, [2025] EWHC 2863 (Ch).
GEMA v. OpenAI and OpenAI Ireland, Landgericht München I (Munich Regional Court I), 42nd Civil Chamber (judgment 11 November 2025).
In re: Zarya of the Dawn, U.S. Copyright Office, Registration No. VAu001480196 (2023).
Tencent v. Yingxun, Nanshan District People's Court, Shenzhen (2019).
Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023).
Copyright, Designs and Patents Act 1988 (UK), s. 9(3) — authorship of computer-generated works.
Thaler v. Commissioner of Patents [2021] FCA 879 (Federal Court of Australia), reversed by Commissioner of Patents v. Thaler [2022] FCAFC 62.
Naruto v. Slater, 888 F.3d 418 (9th Cir. 2018).
Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991).
Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-cv-00613 (D. Del., decided 11 February 2025).
ANI Media Pvt. Ltd. v. Open AI OpCo LLC, I.A. 45300/2024 in CS(COMM) 1028/2024, Delhi High Court, 2026:DHC:5900 (Bansal J., 24 July 2026) — interim order dismissing injunction application.
Comments