12 Nov 6th Annual Symposium on Pop Culture and International Law: Ownership of a Style – Ghibli-Style Generation and the Limits of Copyright within the Territorial Framework
[Lakshmidevi Somanath is Partner – Litigation & Strategy at Anand and Anand, India and a PhD Scholar at IIT Kharagpur. She formerly served in the Intellectual Property Appellate Board, Government of India as a Member Judge.]
In March 2025, the launch of native image generation together with GPT-4o created an avalanche of family photos and movie posters recreated in the unique soft hand-painted style that is specific for Studio Ghibli. The explanation given by the provider defined the line separating persons and organizations, with the model being capable of rejecting a prompt referring to a living artist but accepting those associated with more general studio styles. On the contrary, the copyright law makes no such distinction, and reactions from practicing artists confirmed this fact.
Eighteen months on, the question has moved from the timeline to the docket. In October 2025 the Content Overseas Distribution Association, whose members include Studio Ghibli, asked OpenAI to stop using members’ works for machine learning and to answer infringement claims arising from Sora 2 outputs, and the Japanese Cabinet Office made a formal request of its own. In November, the Regional Court of Munich I held that song lyrics memorised in model parameters, and reproduced in chatbot outputs, infringed German copyright.
The case provides a genuine test for three propositions put forward by clients during a single counseling session, namely: (1) style is unprotectable, (2) training is subject to the laws of the particular jurisdiction in which it occurs, and (3) moral rights arise where economic rights are inadequate. All three propositions hold true, at least in part.
Style Falls Short of the Protection Threshold
First of all, this postulate is well-established and universally accepted. According to TRIPS Article 9(2), copyright covers expression rather than an idea, procedure, or method of operation. Similarly, the Berne Convention only covers the works of literary and artistic nature rather than the manner of their creation. Palette, line weight, weather depiction, and pacing of a cut belong to the latter category and are therefore not protected under Berne Convention.
This position is corroborated in all jurisdictions involved. Japan’s General Understanding on AI and Copyright published by the Cultural Council’s Legal Subcommittee in March 2024 considers the act of infringement at the output stage as involving two elements – resemblance to the concrete expression of a copyrighted work and reliance on the expression. The former is usually presumed if the work in question is included into the training data (as per Tatsuhiro Ueno, who is a member of the subcommittee). Simply being similar to the manner of expression is not enough. In the same fashion, the US Copyright Office has reached this doctrinal conclusion in Part 3 of the report on generative AI training, albeit with some shift in focus – the concept of stylistic substitution is considered as market harm under the fourth factor of fair use doctrine.
The practical consequence is that liability is assessed output by output rather than trend by trend. The forensic question is whether the image is in the style of the catalogue or of a work within it, and the answer rarely turns on the wording of the prompt. Copyright attaches to works and to the rights subsisting in them, and is indifferent to whether an aesthetic is attributable to one director or to several hundred animators.
One Model, Three Training Regimes
Article 30-4 of the Japanese Copyright Act permits the exploitation of a work, by any means, where the use is not directed at enjoyment of the thoughts or sentiments expressed in it, information analysis being the paradigm case. It does not apply where an enjoyment purpose coexists with the analytical one, which the General Understanding illustrates by reference to the collection of works for training intended to generate material similar to those works, and it does not apply where the use would unreasonably prejudice the interests of the rightholder, which captures the reproduction of databases offered on the market for analysis and extends to content that rightholders have sought to withhold by technical means (a fuller treatment appears in the Chambers survey of Japanese practice). Fine-tuning on a single studio’s catalogue in order to reproduce its look is the case at which both provisos are aimed.
The European position is structurally different. Article 4 of the Digital Single Market Directive provides for text and data mining for all purposes with the reservation of the rights explicitly provided that are machine readable in case of content made publicly available via the internet. On top of these provisions, Article 53(1)(c) and (d) of the AI Act requires providers of general-purpose models to introduce a copyright policy that includes identification and compliance with the reservations and publishing a summary of the training content using the template developed by the AI Office in July 2025. These duties started to apply from 2 August 2025, to be enforced from August 2026, along with an extension of a transitional period for existing models (as described by Clifford Chance in its compliance architecture).
A German ruling was not able to provide an answer regarding the training of the model. The EU Court of Justice will deal with the question of training of the models and their output in the case of Like Company v Google Ireland (C-250/25) that has been referred from Budapest in April 2025 and considers whether training and output are acts of reproduction or communication to the public and whether the mining exception applies.
This model could be legally trained in one jurisdiction, exposed in another jurisdiction when the training data is embodied in the weights of the model, and then challenged in all jurisdictions where the output is generated.
Territoriality and the Myth of the Permissive Forum
According to Article 5(2) of the Berne Convention, the scope of protection and the means of recourse are limited by the laws of the jurisdiction where the protection is claimed. Training in the permissive forum does not give any protection against reproduction or communication of work outside the permissive jurisdiction; therefore, the Munich court could assess the activity of the provider regardless of the locus of assembling the dataset.
However, the paradox often discussed in connection with the Ghibli case – the jurisdiction of the home country of the aesthetic field as the most permissive one for training – requires further clarification. The argument that Japanese holders of rights put forward is that prior authorization is required under domestic law, and that an opt-out granted afterwards does not affect the liability for unauthorized training. In light of the provisos to Article 30-4, Japanese law applicable to training of a model targeted at a single rights holder is not so permissive as it is often viewed, and the government is willing to use its powers under the 2025 legislation on the promotion of artificial intelligence if voluntary agreement does not happen.
Pressure also runs in the opposite direction. Recital 106 of the AI Act states that a provider placing a general-purpose model on the Union market should comply with the copyright policy obligation regardless of the jurisdiction in which the copyright-relevant acts underpinning training take place, the stated rationale being that no provider should gain a competitive advantage in the Union by applying lower standards. João Pedro Quintais has argued that it extends the territorial scope of the obligation, and by extension of substantive copyright rules, beyond the provision it purports to support.
Being an advisory rather than a conflict-theory-based dispute, it is not of practical importance. A developer who makes his software for the Union market follows the Union requirement, making the reservation provision in Art. 4(3) the working world default. Thus, the choice of the forum for the purposes of training is a weak tool compared to what one might think; the level of protection depends on the strictest market that the model operates in, and not on the most permissive lab where it was made.
Moral Rights: The Right Claim in the Wrong Hands
Article 6bis of the Berne Convention preserves, independently of economic rights and after their transfer, the author’s right to claim authorship and to object to distortion, mutilation or other modification of the work, or to derogatory action in relation to it, prejudicial to honour or reputation. Japanese law gives effect to this through the rights of attribution and integrity and treats exploitation in a manner prejudicial to the author’s honour or reputation as an infringement of moral rights.
Two structural obstacles stand in the way of deploying that article against style imitation. The first concerns the holder. Moral rights vest in authors personally and, in the civil law tradition, cannot be assigned. The holders are directors, designers, animators and their estates, not the corporate entity that owns the economic rights and that is in practice best placed to litigate. A studio cannot assert what it does not hold, and the authors who do hold it may number in the hundreds for a single film.
Secondly, the matter arises concerning the subject. Integrity means a change in the work. A picture created in an identical fashion does not change any other work. It creates a new work. In situations where Article 6bis is relevant, the range of the article becomes narrower and much more practical – those works that transform protected expression into something that the author can prove to be derogatory and those works that are presented or distributed as coming from the studio.
That narrower reading matches the complaint actually made in 2025, which was not that a palette had been taken but that a name, a reputation and four decades of accumulated goodwill were being used to market someone else’s product. The responsible minister described the works at issue as irreplaceable national treasures. That is a claim about misappropriation and attribution, and the instruments capable of answering it sit largely outside the copyright chapter of the Berne Convention.
Implications for Practice
For the rightsholder, the analysis should shift from inputs to output. It is quite easy to demonstrate that the works have been used in the training dataset, and this shifts the burden of proof to the issue of similarity of expressions. It requires the prompt-output-comparator triad to be preserved simultaneously. Reservations under article 4(3) of the Digital Single Market Directive should be expressed in machine-readable form whether or not the holder is established in the Union, since that is the mechanism the AI Act and the Code of Practice are built around. Training content summaries published under article 53(1)(d) should be monitored, subject to the limitation that the template calls for a general account of sources rather than a work-by-work disclosure. Claims founded on name, mark and attribution should be pleaded alongside copyright rather than after it.
For software creators, the denial system based on the names of alive artists acts as a reputation-related measure rather than one addressing legal liability. According to the decision of the Munich Court, output-based measures aimed at protected individuals and scenes are legally more relevant than any classifier of styles. It is necessary to ensure the appropriate provenance information for a summary according to Article 53(1)(d) right from the moment of ingestion of the material into the database, and not after.
Copyright can determine whether a particular image reproduces protected expression. It cannot determine whether a studio should control the diffusion of a visual grammar that it made legible to the world, and nothing in Berne, in TRIPS or in the instruments built upon them suggests that it was ever intended to. Whether that gap should be filled is a legislative question, and the territorial architecture of the Berne system guarantees that it will be answered differently in Tokyo, in Brussels and in Washington. A borderless tool will have to live with bordered answers.