A Rights Dilemma – Prompts For Music 

Artificial Intelligence (AI) has brought within the realm of reality, what was once a mere figment of imagination – a page from a science fiction novel¹ or a scene from Star Wars! Of course, it is of little surprise that the impact of AI is most palpably felt in the entertainment industry where recent trends have posed before the lawmakers, the classic Collingridge Dilemma – whether to curb innovation in its nascent stages by the imposition of regulations, or to let innovation thrive at the cost of rights of specific sections of society (such as artists and musicians).²  

Speaking of regulations, the Copyright Act, 1957 (“the Act”) is the key legislation that governs and grants protection to authors of original musical work, for the intellectual property in the lyrics, composition and sound recording.  In the larger interest of promoting the creation and publication of original music, this law was intended to prevent original works from being copied, distributed, performed or reproduced without the author being credited and/or duly remunerated for the same.  However, with the recent turn of events, AI is being used (or perhaps, mis-used) to autonomously create musical works.  In a recent case, AI music generators, Udio and Suno were sued by international record labels including Universal Music, alleging copyright infringement in training their models on published music to which the record labels had exclusive rights, without a license to that effect.  

But why is the rights conundrum relevant here?  

AI models are trained on large volumes of existing data and is processed through neural networks into large language models in which it is contextualised. To create a large language model for music, an AI music generator will not only need access to phenomenal volumes of non-linear data (notations, sound recordings, lyrics, styles, voice, language, etc.) but will also consume a large amount of computation power, to generate any original output which will consequently entail heavy costs.  However, if one chooses to train existing large language models on specific and limited data sets, it would definitely be possible to “prompt” the system to generate new musical works, which would be new but not original.  The works so generated in this manner, would most certainly bear significant resemblance and can even be almost identical to songs, styles, and lyrics of popular artists, so much so, that a listener could easily be deceived to believe the song to be another version or derivative of the original work which forms the data set for training the model.  

Can the output really be protected as an independent and original work, or would it qualify as a derivative of the original?  

The Act grants protection in respect of the copyright in any original work only to the author or any person or entity to whom the author has assigned such rights in writing. In respect of a sound recording, it is only the copyright holder who is entitled to create other sound recordings embodying the lyrics or musical composition of the original work including storing of it in any medium by any means. Any other person who creates derivatives of the sound recordings i.e. creates and/ or publishes adaptations, translations or remixes of the sound recordings without a legitimate license from the holder of the copyright in the recordings, would be deemed to be infringing the said copyright, entitling the original holder to not only initiate legal proceedings but also to royalty from the monetisation of such derivatives.  

This leads us to another pertinent question of ethics and bias in the deployment of AI. The spirit of the Indian law gives the highest importance to ethical considerations of artists being attributed for their original work in case of reproduction of such works in any other mode, medium or format. This constitutes the moral rights of the author. Further, the Terms of Use of most publishing platforms carry caveats against unauthorised scraping, violation of which is tantamount to a breach of contract and monetisation of the proceeds may also warrant penal action under applicable law.  

Ethical biases of AI models are further challenged with Generative AI now giving way to Cognitive AI, the technology is trained to perform cognitive tasks in the same way as a natural human. We can therefore reasonably expect the lines of distinction between AI generated output and original work being further diluted in the foreseeable future, thereby making the ethics debate even more complex. An examination of the global judicial landscape – the war waged against Udio and Suno or the recent cease-and-desist notice issued by Tupac Shakur’s estate against Drake for the latter’s release of the song “Taylor Made Freestyle” which was allegedly created using AI and bore significant resemblance to Shakur’s prior work – is indicative of the potential legislative intervention that can be expected in the space. The copyright statutes of most countries, including India, attribute authorship to human creators and not technology. Irrespective of whether the attribution should be to the human who created the work or the machine, the resemblance to existing and copyrighted work cannot be overlooked or undone. Though the decisions of these matters are eagerly awaited by the music fraternity, one can reasonably expect ownership of copyright to be attributed to the first author or assignee thereof and for unauthorised production of derivatives to be challenged in court.   

When speaking of ethical considerations, a common defence taken in matters of data scraping is the “fair use” doctrine – an exception to the mandate for procurement of a license to use copyright work in the event it is being used for “fair dealing” and is not a computer program. Whilst dealing in the work for personal use or for research and educational purposes or for the purpose of reporting, news, or critique should not constitute infringement. Pertinently however, in cases of scraping of published music to generate works using AI that substantially match the original, the claim of fair use would be a stretch. While the matters referred above have been filed in the courts of the US, AI is definitely making the World flat – Indian courts could also be expected to adopt a similar trail of thought.  

Regulation by way of a licensing regime integrating human and AI generated content with specific thresholds seems to be the need of the hour. However given the pace of technological evolution, the task sure does seem like an uphill battle for regulators who are bound by due process even in the drafting of laws.  

But is AI really the Frankenstein it is being portrayed to be – a weapon of mass destruction and obliteration of the human mind? Can it not be leveraged in the interest of creators? Indeed – with new forms of music properties coming to the fore combining cinematography, live performances, competitions and freebies in a package with a sound recording and composition, the rights matrix involves a lot more entities. Even in a project commissioned by a platform or a brand, the rights in the master are assigned whereas the underlying rights may often need to be divided between the artists in the proportion of their contribution. For a human resource to count the number of words in the lyrics or notations in the composition and attribute them to individual artists might be a tedious task. However, with the classification models developed in AI, it can be used to accurately bifurcate different types of data and make the necessary calculations for attribution of rights. Classification models can also be used to bifurcate the deliverable of each stakeholder in the music property (eg. lyrics, editing, composition, performance, etc.) and divide the royalties accordingly configured by way of preset rules emanating from legal provisions.   

Similarly, for ease of acquisition of music properties, document intelligence tools can be taught to scan the contracts of each stakeholder against specific parameters for a quick risk assessment. With music properties becoming high value assets, it is critical to assess the rights being transferred in an acquisition or distribution deal and save the platform or record label a significant amount of money.  

The advent of AI has undoubtedly created a lot of fear and insecurity amidst every section of society. Redundancy of human labour and loss of employment are the core fears underlying all movement against AI and are by no means misplaced. However, let us also acknowledge that the capability of AI is limited to the extent of data currently available as at any given point of time. It does not account for new forms of data that are yet to be generated, at least not with any reasonable accuracy. Therefore, to keep AI alive and retain the novelty of its usage, human creativity must go on.  

[1] Kyet234, “Exploring the Boundless Potential of Home Automation: A Futuristic Integration of Technology and Convenience”, Nov. 29, 2023. Exploring the Boundless Potential of Home Automation: A Futuristic Integration of Technology and Convenience - Kyet234

[2] Krisna Ravi Srinivas, “Two Reasons AI Is Hard To Regulate: the Pacing Problem and the Collingridge Dilemma”, May 02, 2023 and Updated July 4, 2024. Two reasons AI is hard to regulate: the pacing problem and the Collingridge dilemma - The Hindu