This essay was written during a research internship supervised by Nandita Saikia and Sidharth Chopra at Saikrishna and Associates in March 2023.
Music Copyrights in the Age of AI
Khalil Gibran once apparently said: “Music is the language of the spirit. It opens the secret of life bringing peace, abolishing strife.”[1]
However, the use of AI in relation to music has brought with it not peace but a significant amount of confusion. Over the past decade, the use of AI in the arts has developed[2] particularly to generate images possibly from text descriptions,[3] and to make music.[4] In this context, while generating could be understood as a subconscious, automated or mechanical process, ‘creating’ could signify a degree of original thought, artistry, creativity, and intent. For example, a composer ‘creating’ music signifies their involvement and inspiration in the composition of the final work whereas an algorithm ‘generating’ music (either independently or in ‘collaboration’ with a composer) signifies a lack of personal vision, and, possibly, of original expression.
With the integration of AI in producing music, new opportunities for creative expression and innovation have emerged. Since the jurisprudence surrounding AI-generated works is nascent, discussions about the various implications of using AI to create music and the possible ramifications of such use within the context of copyright law are necessarily speculative.
In India, copyright is primarily governed by the Indian Copyright Act, 1957, and copyright protection is granted only to the original[5] or non-infringing[6] expressions of ideas — and not to ideas alone — provided the expressions meet the requirements laid down by the law to be accorded protection.
The copyright statute defines a musical work as one ‘consisting of music’; the definition ‘includes any graphical notation of such work’ within its scope but excludes accompanying words and gestures.[7] Despite this, music is not confined to musical works or to instrumental compositions: it could also include lyrics which are considered to be literary works by copyright law.
Melodies of Machines
Music may be defined as ‘a pattern of sounds made by musical instruments, voices, or computers, or a combination of these, intended to give pleasure to people listening to it’.[1] Lyrics also form a significant part of music, playing a crucial role in conveying the depth and emotion of a song.
Despite the universal nature of music, musical styles around the world exhibit significant differences from one another. While Western music is largely focused on chords and harmony, Indian classical music is primarily melody-based, with a wide range of complex melodic structures known as ragas which are loosely comparable to scales. The use of microtones (which could be considered to lie between the notes of the Western chromatic scale) is also common. Since current AI systems appear to be overwhelmingly based on Western musical styles and structures, these nuances are not necessarily clearly reflected in the music which AI has a role in making.
Nonetheless, Artificial Intelligence has made significant strides in recent years, offering valuable assistance in the creation, production, and curation of music through advanced algorithms and machine learning techniques.
Artificial Intelligence can be used to help compose music by enhancing human creativity and machine intelligence. One of the ways AI enables composers to create music is by allowing them to upload their compositions on the system and create new variations.[2] Another way it aids composers to create new music is by drawing on music theory, orchestration, mood, or genre and generating new music based on existing datasets[3] which are, in turn, based on the notes, dynamics, and articulations in particular pieces of existing music. In this way, AI supports composers to pursue their artistic vision, particularly those who lack the technical know-how entrenched in music theory.
Songwriting is the skill of writing lyrics that fit with the melody and rhythm of music and evoke a particular mood or emotion. Since it requires experimentation, AI can be used to facilitate the process possibly by using it as a starting point to brainstorm ideas based on text prompts relating to the mood and theme of the song.[4] In fact, a lyricist facing writer’s block can use its various customisations and suggestions regarding themes, prompts, syllable structure and genre as inspiration.[5]
Mere composition and songwriting do not generally produce cohesive, marketable music, and bringing music to the public ear requires it to be performed and recorded and then, more often than not, to be ‘processed’, so to speak.
Vocal processing ranges from auto-tuning recordings[6] to using deepfake technology to mimic a person’s voice in a ‘performance’, and involves manipulating vocals through techniques like editing, compression, and sound equalisation to achieve the desired sound effect. In this context, Artificial Intelligence can be used to adjust vocal levels, and automate vocal riding without colouring the vocal track.[7] Combining AI automation with manual adjustments allows the user greater control and flexibility.
Whether or not music includes vocals, the production process facilitates the creation of a finished music recording capturing the intended artistic vision of musicians and includes a wide range of technical and creative tasks such as recording, mixing, and mastering. It enables musicians to create high quality music and distribute it on major streaming platforms while occasionally reducing costs by eliminating the need to engage sound technicians and to rent professional studios.[8] AI’s mixing features can be used to generate custom effects settings based on the sonic palette[9] of a given track[10] making the production process easier.
If recordings of music are distributed through streaming services, those services may use AI to analyse users’ listening habits and recommend new songs and artists to users based on their preferences. Apart from the music itself, these companies rely on the songs’ metadata to train their algorithms to curate personalised playlists based on mood, genre, and popularity, recommending them to users in the hope of increasing their engagement.
Thus, Artificial Intelligence can play a role in every aspect of the development of a commercially viable musical product.
The Intricate Labyrinth of Authorisation
The use of AI in music creation raises questions about the legal ownership of the resulting works and the right to communicate such works to the public.[1] In general, the copyrights subsisting in lyrics are first owned by their authors,[2] those in musical works by their composers,[3] and those in sound recordings by their producers.[4] The musicians who perform music that is recorded also have performers’ rights[5] vest in their performances.
Thus, there is no straightforward single copyright in music, and the rights in recorded music tend of form a complex web involving numerous persons who act at various stages of the composition, performance, recording and distribution of music.
Nonetheless, legally communicating music to the public requires the permission of copyright owners’ either through assignments of ownership or licensing agreements.
For all practical purposes, the only limit to how rights in music can be segregated (with reference to factors such as exclusivity, territory, duration, and purpose) is laid down by the imagination of those involved in determining the scope of grants of rights. Apart from voluntarily negotiated licences,[6] specific conditional rights may also be granted to licensees through compulsory and statutory licences issued under the auspices of the Indian Copyright Act, 1957. Amongst other things, in statutorily specified circumstances, compulsory licences potentially facilitate making hitherto unavailable works available to the public,[7] and help facilitate access to protected works by persons with disabilities[8] while statutory licences may be availed of to make cover versions,[9] and to broadcast music.[10]
Music licences can also be categorised without direct reference to the copyright statute. A master licence, for example, generally permits the licensee to use the sound recording of a protected song; a mechanical licence to reproduce and release the audio of a song usually in physical form; a sync licence to integrate the audio of music with visuals; a print licence to reprint musical notes or lyrics; a theatrical or grand rights licence to perform music in a live performance of a play; and a public performance licence to perform and broadcast music beyond the confines of the theatre.
Unsurprisingly, rights in music are often contested, and the licensing process, complex at the best of times, can present challenges for works generated using AI or composed with the assistance of AI, primarily due to uncertainty surrounding their ownership.
Composite works, such as music, involve multiple stakeholders, including composers, lyricists, performers, producers, and record labels, making it particularly challenging to determine issues relating to their authorship and ownership. Introducing AI into the equation significantly muddies waters — traditional copyright law is not geared to contend with it, since the law does not favour recognising authorship by AI and AI companies may attempt to contractually determine questions of musical ownership in their own favour. Due to this, to promote licensing and mitigate the risk of copyright infringement, it has become essential to establish clear guidelines for identifying the authors and owners of works in whose creation AI has played a role.
Algorithmic Artistry or Copyright Infringement?
AI systems and models rely on large datasets, such as lyrics, audio compositions, and texts, images and videos, which are processed by complex neural networks to generate novel material. As the system is continually trained with more data and human feedback,[1] it improves. Current AI systems are overwhelmingly based on Machine Learning which uses data and algorithms to imitate the way that humans learn with gradually improved accuracy and effectiveness.
While using AI-Assistive tools to enhance the creative process is unlikely to attract claims of infringement when musical output is created, the use of generative AI-models presents a challenge in terms of avoiding copyright infringement during each step of the process of using AI to make music.
It is crucial to consider the datasets used for training AI when utilising AI models for composition and songwriting. These datasets may contain copyrighted material, and due caution must be taken to avoid infringement ideally by ensuring that the music included in datasets is either appropriately licensed or in the public domain.
Section 52(1)(a) of the Indian Copyright Act, 1957, provides for a Fair Dealing exception to copyright infringement, which permits the use of copyrighted material for specific limited purposes without authorisation from the relevant copyright owner. Determinations as to whether the dealing with protected works is indeed ‘fair’ may be made with reference to various factors such as the amount and significance of the work used, the intended purpose, and the likelihood of competition between the two works, the work which has been used without authorisation and the work in which it is used. This approach is akin to the four-factor test of Fair Use stipulated in Title 17 US Code Section 107.
Given the lack of legal precedent regarding the use of AI in creative works, courts may look to the 2015 decision of the US Court of Appeals (2nd Circuit) in Authors Guild Inc. v. Google Inc.,[2] even though foreign judgments have only persuasive value in Indian courts. In this case, the digitisation of copyrighted books to facilitate search functionality with snippets from the books being displayed to users was deemed fair use due to the transformative nature of the work; the ‘revelations of parts of the originals’ were determined not to ‘provide a significant market substitute for the protected aspects of the originals’.[3]
By analogy, in relation to creating music, it is possible to argue that the deployment of generative AI models which create new works without substituting original works does not give rise to liability for copyright infringement due to the transformative nature of the outputs generated. On the same basis, in relation to distributing music, it should also be possible to argue that the storage of a song’s metadata by streaming services for recommendation purposes should not be considered copyright infringement since it does not act as a replacement for the original work.
However, while AI has been beneficial to enhance and develop some aspects of artistic expression and access to the same, it has also presented challenges.
Once instrumental music has been generated, for example, if it is to be accompanied by vocals, employing Artificial Intelligence to process the vocals and aid in the production process can be cost-effective. However, caution must be exercised when using AI: vocal processing involves modifying vocals. While even workaday vocal processing can potentially result in the moral rights of a performer being infringed, in the case of simulating a person’s voice with deepfakes, the supposed performer’s right of publicity (which enables them to control the use of their likeness especially in commercial contexts) could be violated not to mention that just the use of a deepfake could, depending on the circumstances, veer towards the commission of fraud.
Currently, the law is unclear. It is therefore imperative to revisit the existing legal framework and make necessary amendments to squarely address the role of AI in the music industry.
Developing Regulatory Regimes
Technological developments necessitate corresponding changes in law. In India, the 2021 Rajya Sabha Standing Committee Report[1] recognised the benefits and impact of AI; from a policy perspective, the creation of separate categories of rights for AI and AI-related inventions and works may be recommended.
Other countries, too, have been exploring how best to recognise and regulate the proliferation of AI. In March 2023, the US Copyright Office launched an initiative to examine the various copyright issues raised by the use of AI.[2] This included a new registration guidance[3] stipulating the disclosure of inclusion of AI-generated content in registrable works, and organising listening sessions to discuss the utilisation and implications of use of AI in the creative fields.[4]
The regulatory framework should be developed keeping in mind the rationale behind granting copyright protection to works. The labour theory of copyright grounded in Locke’s theory of property suggests that a person investing time, money, knowledge, and other resources in creating intellectual property should receive the ownership rights over it on account of their labour.[5] The welfare theory, on the other hand, promotes the interests of society as a whole and favours the greatest good for the greatest number of people.[6] It aims to balance incentivising creators with mechanisms created to make works available to the larger public.
In order to reconcile the opposing views, considering the growing presence of Artificial Intelligence in the creative realm, it is essential to address the issue of authorship and ownership of AI-generated works. In works created by Human-AI collaborations, where AI serves as an aid in composition, copyright may belong to the human seeing as the human is likely to provide the creative input with the AI merely providing technical assistance.
As for generative AIs, depending on the nature of the AI system, the copyright may belong to the developer or to the individual who contributes input to the AI or to the person who prompts the AI to generate output. While current AI systems are limited to specific functions, designating AI systems as co-authors of a work could perhaps encourage the development of more advanced AI models capable of independent action. This would help clarify the legal framework surrounding AI-generated works and foster innovation and creativity while ensuring appropriate attribution and recognition for all parties involved.
Granting copyright protection to AI-generated works would presumably incentivise software developers to create new and advanced technologies which could further enhance musical creativity, enabling musicians to refine and develop their art. The integration of AI in music production could provide new avenues for creative expression while human artistry adds emotional depth and nuance to the final product, paving the way for the emergence of a more innovative and dynamic music industry.
Moreover, existing generative AI algorithms are capable of analysing and learning from both copyrighted and non-copyrighted materials. To recognise the welfare theory, it may be useful to expand the current fair dealing exceptions to copyright infringement to include data mining and machine learning exceptions, as Japan could perhaps be considered to have essentially done.[7] This could enable AI to promote public welfare by allowing greater access to information. Furthermore, it is important to establish liability for the infringement of rights in AI-generated works to uphold the labour theory of copyright by ensuring that creators of AI-generated works are properly recognised and compensated for their efforts.
Instead of merely viewing Artificial Intelligence as an obstacle to artistic expression, embracing and promoting AI-Human collaborations to create music, and developing a clear regulatory regime to govern them, may assist developers in ensuring that future AI models are developed while taking copyright limitations into account. As AI is a growing field, the legal framework must be designed with the constantly evolving AI landscape and its capabilities in mind, without compromising creativity and originality.
[1] Khalil Gibran, available at: https://www.goodreads.com/author/quotes/6466154.Kahlil_Gibran?page=2
[2] References to companies in this text are intended to be no more than references to companies whose work in the field of AI it may be worth considering in relation to the issues under discussion. They are not a comment on either any company or any AI system, programme, or algorithm.
[3] DALL-E: Creating images from text – OpenAI, available at: https://openai.com/research/dall-e
[4] Magenta, available at: https://magenta.tensorflow.org/
[5] Literary works
[6] Sound recordings and Cinematograph films
[7] Section 2(p), Copyright Act, 1957 (Act 14 of 1957)
[8] Music, Cambridge Dictionary, available at: https://dictionary.cambridge.org/dictionary/english/music
[9] AIVA Technologies, available at: https://www.aiva.ai/about
[10] Amper Music, available at: https://www.shutterstock.com/music/search?artist=amper-music
[11] OpenAI: Chat-GPT, available at: https://chat.openai.com/auth/login
[12] LyricStudio, available at: https://lyricstudio.net/?via=marco
[13] Antares Auto-Tune, available at: https://www.antarestech.com/
[14] Vocal Rider, available at: https://www.waves.com/plugins/vocal-rider#how-to-level-vocals-perfectly-in-3-clicks
[15] LANDR, available at: https://www.landr.com/
[16] Sonic palette is a MIDI controller type instrument which does not create musical sounds by itself, instead it generates data which is interpreted by a synthesiser.
[17] iZotope, available at: https://www.izotope.com/ Also: MelodyML, available at: https://melody.ml/
[18] Section 2(ff) “communication to the public” means making any work available for being seen or heard or otherwise enjoyed by the public directly or by any means of display or diffusion other than by issuing copies of such work regardless of whether any member of the public sees, hears or
otherwise enjoys the work so made available, Copyright Act, 1957 (Act 14 of 1957)
[19] Section 17 read with Section 2(d)(i), Copyright Act, 1957 (Act 14 of 1957)
[20] Section 17 read with Section 2(d)(ii), Copyright Act, 1957 (Act 14 of 1957)
[21] Section 17 read with Section 2(d)(v), Copyright Act, 1957 (Act 14 of 1957)
[22] Section 38, 38A and 38B, Copyright Act, 1957 (Act 14 of 1957)
[23] Section 30 and 30A, Copyright Act, 1957 (Act 14 of 1957)
[24] Section 31 and 31A, Copyright Act, 1957 (Act 14 of 1957)
[25] Section 31B, Copyright Act, 1957 (Act 14 of 1957)
[26] Section 31C, Copyright Act, 1957 (Act 14 of 1957)
[27] Section 31D, Copyright Act, 1957 (Act 14 of 1957)
[28] Reinforcement Learning from Human Feedback (RLHF) is a technique that trains AI models from human feedback to predict whether an output is high reward (good) or low reward (bad).
[29] Authors’ Guild v. Google Inc., 804 F.3d 202 (2d Cir. 2015)
[30] Id. at 229.
[31] Review of the Intellectual Property Rights Regime in India, Department Related Parliamentary Standing Committee on Commerce, available at: https://rajyasabha.nic.in/rsnew/Committee_site/Committee_File/ReportFile/13/141/161_2021_7_15.pdf
[32] NewsNet Issue No. 1004, Copyright Office Launches New Artificial Intelligence Initiative, US Copyright Office, March 16, 2023, available at: https://www.copyright.gov/newsnet/2023/1004.html
[33] Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence Available at: https://www.copyright.gov/ai/ai_policy_guidance.pdf
[34] See more at: https://www.copyright.gov/ai/
[35] Janhavi KM, Theories of Intellectual Property Rights, IP Matters, January 19, 2021, available at: https://www.theipmatters.com/post/theories-of-intellectual-property-rights
[36] Jessica Meindertsma, Theories of copyright, Copyright Corner, Ohio State University, May 9, 2014, available at: https://library.osu.edu/site/copyright/2014/05/09/theories-of-copyright/
[37] Article 47 states that use of copyrighted work for the purpose of machine learning will be allowed without the authorisation of copyright holders. Copyright Law of Japan, available at: https://www.cric.or.jp/english/clj/cl2.html#art47-5
