First published by AsiaIP.
Author: Lakshmidevi Somanath
Copyright law has traditionally addressed the asymmetry between the costs of creating a work and the comparatively low marginal costs of reproducing it. By granting authors exclusive rights for a limited period, the law seeks to preserve incentives for creative production while balancing those incentives against public access, dissemination and subsequent innovation. The emergence of generative artificial intelligence places this balance under renewed pressure, not merely because reproduction has become easier, but because the processes of creation, commercial exploitation and substitution are increasingly automated.
This difference is demonstrated in the Indian gaming market. India was home to between 515 and 525 million gamers in 2025, which made it the biggest gaming market in the world.[1] The commercial value of a game consists not only in the software itself, but also in the characters, visual worlds, narrative, music, and unique artistic components which give the game uniqueness and commercial value. In light of Indian studios increasing their efforts to produce games intended for both domestic and foreign markets, the issue of the legal protection of such components becomes relevant. Generative AI raises concerns in this regard because on one hand, it makes the production of such content cheaper, but on the other hand, it makes it more difficult to avoid reproducing elements that are recognizable in already existing franchises.
Within the video games sector, the character design, the visuals, artwork, the sound, music, dialogue and other aspects of the narrative may be considered independent protectable subjects. However, their protection is conditional upon identifying the protected expression instead of the idea behind, the method of operation or even the game mechanics used. This issue becomes highly relevant in case when game studios resort to using AI to create or edit game assets. In such cases, although the use of the AI model itself is lawful, its output may reproduce copyright protected expression. Studios that create game content through artificial intelligence may be at risk even if the training for these systems was perfectly legal.
In July 2026 Deezer reported that more than half the music arriving on its platform each day was made by machines. That is roughly 90,000 tracks every twenty-four hours. The company said that now that half of all daily uploads are AI-generated tracks, they were taking additional steps to protect artists and songwriters.[2] Fifteen months earlier the figure had been 18 per cent.[3] It has been stated by Deezer that as much as 85 per cent of streams related to tracks created using artificial intelligence technology in the year 2025 were found to be fraudulent. It should be noted that this number pertains to streams for tracks created using artificial intelligence and not the streaming service in general.[4] Much of this material comes from streaming farms, which flood a catalogue with machine-made tracks and then fake the plays to draw money out of the royalty pool.
India is one such country where the economic reliance of creators on digital music platforms can be seen very distinctly. As per the Annual Report of the Indian Performing Right Society (IPRS) for the fiscal year of 2024-25, the society collects ₹741.6 crore and distributes ₹608.8 crore in royalties annually.[5] Streaming revenues contribute about ₹602.4 crore to IPRS, constituting over 80 per cent of the society’s overall collections. It is clearly evident from the above figures how integral digital platforms have become to the payments that creators of music in India receive. But where streaming dominates the royalty collection process, any form of manipulation of streaming statistics, the creation of fake plays and the replacement of human-made music with artificial music become issues.
Global recorded music revenue reached 31.7 billion dollars in 2025, up 6.4 per cent on the year before.[6] According to a study carried out on behalf of CISAC, it is projected that by the year 2028, up to 24 percent of revenues of music rightsholders and 21 percent of revenues of audiovisual creators will be at risk from the use of generative AI. The study forecasts total revenue losses of up to €22 billion for a period of five years.[7]
Generally speaking, when a song is commercially released, there are distinct rights over the musical composition and lyrics, as well as the sound recording capturing the work. Thus, a streaming platform can need both licenses in relation to a single piece of music, since the rights to the sound recording and the musical work may be exercised independently by different parties. Such rights may be managed by distinct copyright holders and licensing schemes, which means that the commercial use of the track is possible only through the cooperation of several parties and their respective rights. In India, according to Section 14 of the Copyright Act, 1957, exclusive rights include the rights to communicate the work and the sound recording to the public.[8] The popularity of streaming platforms has brought additional complexity and number of transactions into the licensing process, but identification and proper crediting of the creators still depend on the use of the metadata. The issue of AI-generated music also brings about the question of whether the track created by AI could be found infringing, and whether it could compete for the same pool of royalties with human-generated content.
Inputs and Outputs
There are two distinct issues relating to copyright law that must be analyzed separately in cases of copyright infringement by generative AI. The first relates to the inputs in terms of whether reproducing protected works to develop an AI model itself amounts to copyright infringement, or falls within an exception. The other issue relates to the outputs, which would involve determining whether any outputs produced by the model amount to reproduction of a substantial portion of copyrighted work or constitutes an infringement of a copyright.
The decisions made by the US courts exemplify the importance of being fact-based when making a determination of fair use. In June 2025, in Bartz v. Anthropic, the Northern District of California determined that using lawfully obtained books for training a language model was transformative and amounted to fair use.[9] Nevertheless, the court went ahead and distinguished the training of the language models using lawfully obtained books from the storage of pirated versions of the books in the central library of Anthropic. In the case of Kadrey v. Meta Platforms, the court found in favor of Meta Platforms with regard to the determination of fair use, but recognized the ability of generative AI to dilute the market through flooding it with similar works. The court also recognized that the plaintiffs had failed to make the necessary showings in terms of market substitution.[10]
Similarly, the difference between inputs and outputs is evident in the lawsuit that Disney and Universal have filed against Midjourney in June 2025, and subsequently by Warner Bros. Discovery in September 2025.[11] This litigation raises concerns, not just regarding the creation of images that depict copyrighted characters even after minimal prompt, but also regarding the use of copyrighted works in the process of creating and training these technologies. The case of Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence in September 2026 at the Third Circuit found that ROSS’s use of the protected headnotes in developing their competing legal research system was not a fair use. The court acknowledged the existence of the legal research market as well as the licensing market for the copyright material used in creating AI software. However, it should be carefully examined in light of its distinction from generative cases in the field of AI.[12]
The English High Court in Getty Images v Stability AI dismissed Getty’s claim of secondary copyright infringement, holding that the model weights did not qualify as an infringing copy of the copyright work that was being alleged to have been infringed for the purpose of the claim. This judgment is of importance for the consideration that the court took on the statutory criteria for determining secondary copyright infringement and the connection of model weights with the storage of copyright works. However, the ruling cannot be taken as definitive for the fact that AI models will never include an infringing copy.[13]
In India, the Copyright Act, 1957 does not specify that video games belong to any particular class of protected works. Instead, protection needs to be approached on the basis of different classes of copyrights under which the individual components of a game might fall, such as computer programs, artistic works, literary works, musical works, and sound recordings. This approach to component-based protection has implications for AI-aided game development. While the developer has clear right to its own software and creations, such right does not necessarily transfer to third-party characters or artistic elements included in the AI-generated material. Moreover, while the use of an AI model for development might be lawful, this alone would not mean that the resulting asset does not violate somebody else’s copyright.
Within the EU, Article 53 of Regulation (EU) 2024/1689 mandates that the providers of general-purpose AI models have a policy for their compliance with the EU Copyright Act, as well as procedures for identifying and honoring the rights reservation under Article 4(3) of the DSM Directive and publishing sufficient information about the data used for training the model. Such obligations, are not a mandate for disclosing the full training dataset of AI models.[14] The Australian Productivity Commission assessed the viability of an exception for AI training through text and data mining under their review of copyright and digital technology policies. The Australian Government later declared in October 2025 that they will not implement this exception. This is due to a decision made on policy grounds that the current copyright policy be maintained rather than creating an exception for AI training.[15]
The direction of reforms
In December 2025, the DPIIT released the first part of a working paper on whether the provisions of the Copyright Act, 1957 adequately address the usage of copyright protected materials for AI training.[16] The committee recommended that a mixed model should be adopted where developers of AI systems should be granted a blanket license to use copyrighted material lawfully obtained for their training and development, with royalties payable once the system becomes commercially viable. In the proposed mixed model, royalties shall be set by a government appointed committee. The proposed model acknowledges two competing institutional realities: negotiations between the parties might be too costly, while lack of compensation might shift the economic value of creative labour to technology companies. Nevertheless, the proposal raises important concerns about how royalty rates will be set, how rights owners in different categories will be represented, how works without an owner will be handled and how royalties will be distributed among individual creators. This particular issue is especially relevant to the Indian music and video games industry where ownership might be shared by many: composers, performers, publishers, developers and production studios.
The US Copyright Office is concerned about replacing market licensing with a statutory one. Meanwhile, bilateral licensing is less likely to be convenient for individual creators and performers and small studios rather than large record labels. Collective licensing might solve the problem. A collective management organization that represents the interests of the rights owners can arrange licenses, develop remuneration schemes and ensure mass licensing. Yet the legitimacy of this solution can be achieved only through effective representation, proper distribution of revenues and transparency of decision making processes while providing safeguards for those right owners that do not want to participate in collective negotiations.[17]
To view all formatting for this article (eg, tables, footnotes), please access the original here.
etc. whilst wrongfully claiming to be part of our firm and making false claims and allegations.