The case demonstrates the significance ofcontinuously seeking new tools and techniques to enhance musicpublishing. Practitioners can benefit from Diana’s approach of experi-menting with different AI tools and incorporating them selectively tomaintain a unique sound and efficient business operations.

connection in music, even when integrating advanced technologieslike AI. Students should understand that while AI can enhance effi-ciency, the authenticity and originality of content are crucial for audi-ence engagement (Wei et al., 2022).2. Strategic Use of AI: Diana’s selective use of AI tools for specific tasks,such as data management and analytics, illustrates the importance ofstrategically adopting technology to enhance workflow without com-promising creative integrity. This teaches students to critically evaluatethe tools they use and ensure they align with their creative goals(Hampe & Schwabe 2001).3. Adaptability and Continuous Learning: Diana’s approach high-lights the need for continuous learning and adaptability in therapidly evolving field of music publishing. Students should beprepared to stay updated with the latest technological advancementsand be willing to experiment and learn from their experiences

The evolving capabilities of AI in natural language pro-cessing and automated content generation might open new avenues forinnovative music marketing and fan engagement strategies. Diana plans tointegrate AI more deeply into her operations through advanced tools forpredictive analytics and automated content management systems, increasingefficiency and scaling her operations to reach a global audience effectively.

Ensuring that AI does not compromise the emotional authenticity of musicremains a critical challenge, and Diana emphasizes the importance of usingAI to support rather than overshadow human creativity.

Diana faced challenges in aligning AI toolswith existing workflows and ensuring compatibility. However, throughtargeted training and continuous learning, these challenges were addressed.The initial investment in AI systems was a hurdle, but the long-term benefitsof increased efficiency and reduced errors justified the expenditure.

Diana’s success with AI tools under-scores the importance of balancing technological advancements with main-taining the emotional and artistic integrity of music. Her work serves as anexample of how AI can be used to support rather than replace humancreativity in the music industry

providing adequate trainingand resources is essential for fostering intrinsic motivation and ensuring thatusers feel competent and capable of using new tools. This approach notonly enhances proficiency but also builds confidence in using AI technolo-gies (Ezinwa et al. 2024). The continuous upgrading of IT infrastructure tosupport the latest AI technologies is another crucial aspect of developingthe ability to use AI tools effectively.

AI’s capacity to provide real-time analytics aligns with the findings ofWatson and Leyshon (2022), who emphasize the importance of timely andaccurate data in the music industry. Real-time analytics allow for agiledecision- making, enabling music publishers like Diana to adjust their strate-gies based on current market performance and trends. This capability isparticularly valuable in a fast-paced industry where timing can significantlyimpact the success of promotional campaigns.

AI tools have opened up significant opportunities for Diana, particularly indata management and analytics. These tools enable her to access real-timeanalytics on how songs are performing globally, allowing her to make quickdecisions on promotional strategies. AI also provides the capability tomanage artists’ portfolios and rights across different countries and legalenvironments seamlessly

Diana highlighted her motivation by saying:My primary driver is the desire to streamline the complex pro-cesses of music publishing and rights management to ensure thatartists receive their fair share of earnings. AI tools enable us toautomate many of the labour-intensive tasks involved, from track-ing song plays across digital platforms to managing royalties andrights distributions efficiently.

AI-driven financial tools provide significant benefits in terms of optimiz-ing revenue streams and ensuring timely royalty payments (Barata &Coelho, 2021). The ability to analyze market trends and consumer prefer-ences using AI aligns with Diana’s strategies for maximizing revenue andenhancing the commercial success of her music projects

Boomplay’s analytics features have providedinsights into popular songs, influencing the release strategy for albums andsingles. By leveraging AI tools for data analysis, Diana ensures that hermarketing efforts are targeted and effective, resulting in higher streamingnumbers and increased revenue.

On the publishing front, AI tools help track music plays acrossdifferent platforms, ensuring accurate royalty collection and distribution.AI- driven analytics have also provided predictive insights that guide produc-tion decisions, enhancing the commercial success of her projects

AI tools have significantly impacted Diana’s music production and publish-ing processes. A pivotal moment was the training session conducted byVladimir Philippov, CEO of Heaven 11, which enhanced her team’s capabili-ties in using AI tailored for the music industry. The Heaven 11 platformoffers features for music distribution and rights management, crucial forartists aiming to earn from their creativity and have their music distributedglobally

The adoption of AI tools in Diana’s operations reflects broader trends inthe music industry, where AI is used to enhance various business processes.AI technologies like those used by Diana streamline administrative tasksand financial management, which are crucial for maintaining operationalefficiency. This trend is consistent with findings in Chapter 2, where theimpact of AI on the music industry is discussed extensively in a literaturereview.

Diana Hopeson integrates AI tools extensively in her operations as both anartist and a music publisher. She utilizes Google’s business tools for manag-ing interactions and tracking engagements, which streamlines administrativetasks and enhances efficiency. The Oze app is pivotal for her accountingprocesses, simplifying financial transactions and client interactions.

She also attended a needs assess-ment class at the Ghana Institute of Management and Public Administrationand graduated with a master’s degree in philosophy from the University ofEducation, Winneba, in 2019. Diana’s dedication to her craft and continuouslearning have been pivotal in her journey, allowing her to adapt to theevolving music industry and maintain her relevance over the decades.

In March 2021, she was named among the Top30 Most Influential Women in Music by the 3Music Awards Women’sBrunch.In addition to her music career, Diana has played an instrumental role inshaping the music industry in Ghana. She served as the president of theMusicians Union of Ghana (MUSIGA

Diana Hopeson, also known as Diana Akiwumi, has been a significantfigure in the Ghanaian music industry since 1991. With 11 albums and 15singles released, she has made a name for herself as a gospel artist and apioneering music publisher.

Generative AI creation can be both infringing as well as non-infringing, copyright within its periphery has means to deal with the infringing content, copyright should not be used as a tool to end the generative AI creation.69 Copyright law should be used as a tool to correct the infringing aspects, if any in copyright law from time to time

different standards for machine creation. It would also reset the copyright infringement standards to a difficult level that would deter the objective of the Law where one would not be able to create without infringing the Law. The law would become counterproductive and supress creativity instead of encouraging it.

This paper is rather an attempt to say that the infringement analysis should be done on the output of the generative AI and not on its training.

It has explained the major concerns around the concept of training what Prof. Lemley calls as ‘fair learning’. And further has raised questions regarding the infringement analysis of AI-generated music. It has also shown the possible alternate to the existing test of copyright infringement, and the problems of utilising AI tools to address the issues caused by Generative AI.

SUNO AI claims to be the first Author of music when generated using its basic version and if the song is generated through a premium version of the same platform, the user becomes the owner of the song and enjoys a licence to commercially license the same

Even musicians have begun to adopt the usage of SUNO AI, which shows the acceptance of this from the creative community.

having a copyright protection over a work does not guarantee any monetary returns in it, and the vice versa is equally true. Copyright protection is an incentive to create music, it is not a pre-requisite for creating music.

more dissimilarities than similarities.

Though, it may have similarities on the surface, deep within they are different, and cannot be considered as same just because they belong to the same tune.61 It is not possible to grant protection to the tune itself. If given, it runs the risk of monopolising a commonly held substance that is free for all to use.

Within a song there are always ‘protectable’ and ‘unprotectable elements’, and as mentioned in the previous paragraph, melody is the protectable element in copyright law; by replacing the existing test of substantial similarity from lay observers, we might run the risk of granting protection to unprotectable elements and could turn the copyright infringement jurisprudence upside down.

Unless the Generative AI platform generates songs on a particular style/particular voice, it may be difficult to prove before the court that the platform is trained upon the specific copyrighted material. The current form of copyright Act does not look at copyright infringement through the input, it looks at it through the lens of output.

This note-by-note fragmentation has not been the objective of law, if the test were to be replaced with finer tests, one could never create music without violating copyright law, applicable both for humans and machines as the percentage of similarity between two songs could be similar that does not necessarily mean that the songs are substantially similar/same.

The platform does not thieve on any individual's voice or does not use any individual's voice or style, for instance, even if given a prompt to provide music like Mr. Frank Sinatra, it does not provide any output, the idea behind this is probably to exempt themselves from any violation of personality rights/moral rights.

The AI-generated song does not have a separate composition alone. It is created as a recording itself, unlike the traditional understanding of creation, where the composition composed and then it is recorded.

A decade ago, sampling of music raised a similar copyright concern from the right holders. Sampling is a process by which composers rely on portions from a previous song to create a subsequent work, courts in the United States held, sampling of music cannot be done without taking license from the rights holders.

The melody has multiple musical elements that may be working together, the individual elements do not enjoy separate protection under the realm of copyright law. The unprotected element and the protected elements are so intertwined that the protected elements under the musical creation cannot be heard without the unprotected elements.

There is a problem when someone begins to understand the concept of ‘musical copyright infringement’, there is no single accepted definition of a ‘song/musical work’. The inherent nature of ‘music’ makes it difficult to identify the infringing element in the work.

‘Melody’ is the only part of the music that could be notated on a piece of paper, and rest can only be recorded. A song enjoys two separate sets of copyrighted protection, (i) the composition and (ii) recording.

It is impossible to create anything new which does not remotely resemble the previous creation, it applies for creation in general, but it is truer for musical creation in particular.

It is pertinent to remember that human authors also have access to the previous works, and their artistic taste and creation is also possible because of the training that they received from the existing works. According to Professor Mark Lemley, if the AI platform procures access to the work only for the purpose of training it should be exempted as fair learning.

‘training of the AI platform’ is akin to training of humans and that does not per se amount to infringement and it is covered by ‘fair use doctrine’. It is interesting that fair use doctrine has played a key role in copyright infringement suit but has seldom played any concrete role in music copyright infringement cases

If the generated work resembles a particular musical work, then the platform is liable for copyright infringement if it had not taken license. In absence of it, the platform might have to take license from all copyrighted material before training which may not be technically feasible considering the mammoth data that is used for training purpose would involve several stake holders and right holders whose permission would become relevant.

SUNO AI has certain technological protection measures in place that is does not allow the user to generate music like creators, that is, in the same style or tone. This is a commendable move that safeguards the platform against the violation of personality rights and copyright implications.

It is pertinent to mention that the AI platform can replicate the work in the style of a particular artist (may it be writing/composing etc), this could not have been possible unless and until the platform is specifically trained on the data.40

as it produces music without involving any cost of production, at least at present, it doesn't involve any cost.

Generative AIs have made it possible to create a world of increased creation based on the existing body of work,38 in that sense it has transformed the way creativity and art have been perceived since art's inception.

It makes it almost impossible for anyone to predict the outcome of the Generative AI creation and it adds layers of complexity to the creation by itself. Once the outcome is received at the users' end the AI receives feedback on its creation, this helps the AI in decision making and it learns from the mistake made in the past.3

The platform does not create in a human sense with the final goal in mind. Hence, the whole gamut of creation would also differ because of the prompt given by the user.

quite like the training undertaken by a musical aspirant, every creator creates only upon the existing knowledge that they possess in the art form, the knowledge acts as the fodder that enabled the creator, bereft of the knowledge, the creator would not be equipped to create. Even phenomenal music composers like Mozart and Beethoven were trained in music before they began to create music. No creation springs from a vacuum or without existing knowledge.

If the AI is trained on fewer materials, it would hamper the quality of the output and at the same time, it may enhance the chance of copyright infringement.

With the patterns that it has found, it creates a work based on the input it received from the user, who may ask for a particular style or genre of music. To find a pattern and produce a creative output that matched with the input the AI must be trained on thousands, millions or even billions of data.28 Massive amount of these audio files is freely available on various platforms such as YouTube, Spotify and i tunes.29 SUNO AI is trained on all these materials from the internet and the audio files are perceived by it as data.

It could be both copyrighted and non-copyrighted content, these contents are then stored in a numerical form which is then processed by a computer which is later used for training the platform.21

Supervised learning is limited to selecting the exact data on which the model would be trained on with a clear goal of creation, unsupervised learning is when the model is trained on material and is expected to find a pattern and involve in creation, the reinforcement learning is akin to a trial and error where the platform is given specific feedback.

It could be in the form of invention of the micro phone device, the invention of music sampling, tune generator for composer17 or even the algorithmic sequencing of the playlist of the listener18 is also a form of AI. However, Generative AI is something different from the past as earlier AI aided the user in creativity, and merely acted as a staff, today the system has become an independent creator by itself, it is no longer a mere tool.

it has also managed huge influx of funding for the AI creation.14

It is pertinent to bear in mind that these platforms ‘train’ their Generative AI only to create future works, they do not replicate the works as such.12

That too the creation can be done at the click of a button by giving few prompts. In that sense, it has enabled Authors to create without having any pre-requisite knowledge in art, and creation.

Copyright had not been phenomenally successful in staving off any technological development in the past, it has at best worked its way around the technology in the process it has also bridged the long gap that existed between the creator and the consumer. The technological development has also helped the consumer to also become a creator,

copyright law itself is a result of technology. It burgeoned only after the invention of the printing press- a technological development.3 It was introduced to censor the books that were sold in England and to regulate the book trade by governing the relationship between the publisher and the work by providing limited ownership to the publisher for reaping profits from the work.4

Copyright law stands on basic foundations such as ‘individual Authorship, and promotion of creativity.2 These foundation principles have held the sanctity of copyright law since its inception, the validity of these peremptory norms has been currently challenged by the Generative AI creation.

Music is perceived as a process by the human creator, it is perceived as product by the Copyright law and is perceived as data by the machine.

The Authors of this paper have analysed the quandary by creating ‘music’ using SUNO AI, and by performing infringement analysis of the created song through another AI platform MIPPIA to understand the complex terrain of infringement analysis of generative AI.

Platforms like SUNO AI have enabled even non-musicians to create music and don the hats of composers by giving few prompts without understanding the language in the first place This has disrupted copyright's traditional understanding of music and infringement.

As we can see,increasing the size of a model and the amount of training ex-amples increases its quality, but also the energetic cost.

To that end, we proposedthe use of a new evaluation based on Pareto optimality to givean equivalent importance to both model quality and their en-ergy consumption. This places computational complexity andresources at the heart of the research process. It should benoted that our approach is generic and could be applied toany type of model or input data.

his is the first study on energyconsumption for waveform generation and a primer attemptto include energy efficiency in the entire evaluation proce-dure.

Mean OpinionScore (MOS). It is a subjective measure ranging from 1 to5, based on a qualitative test where participants are askedto rate as 1 the lowest perceived quality and 5 the highestwhen comparing a set of results

it consumes 64.8 kWh, which is slightly higher thanour initial estimation (certainly due to CPU and DRAM en-ergy draw). At this point, we believe that comparing modelspurely on the basis of these estimations is questionable, andargue that the real energy should have been recorded.

Approximated energy consumption for training sev-eral state-of-art generative audio models.

first, an auto-encoder istrained for 12 hours, then a sequence generator for 10 hoursand finally an end-to-end fine-tuning for 30 hours. All partsare trained on 4 NVIDIA P100 GPU

These include the hardware usedto train the model, such as the type of GPU and total trainingtime in hours. Surprisingly, we found out that only five ofthe studies properly specified both criteria.

the totalnumber of model parameters

”Carbontracker”2, which tracks and predictsenergy consumption and carbon emissions for training deepleaning models. This provides a more accurate estimationwhile being user-friendly.

the amount of energy required to train a model(until convergence), and the amount of energy required bythe model for inference (generating a sample in the case ofaudio synthesis)

However,the generated samples tend to be slightly blurry compared torecent adversarial networks, such as WaveGan [10] or GAN-Synth [11]. These show impressive reconstruction abilitiesbut lack latent expressivity and are difficult to optimize dueto unstable training dynamics.

Furthermore,they also provide almost no direct control on the generativeprocess.

In red (left) those thatrefer to the quality of the generated samples, and in green(right) those that refer to their algorithmic complexity andperformances.

Given trainingdata points x following an unknown probability distributionp(x), generative models aim to learn a parametric distribu-tion pθ (x) that approximates p(x), by iteratively changingmodel parameters θ

hen,we propose the use of a multi-objective Pareto optimality cri-terion to provide fair comparisons simultaneously on genera-tion quality and energy efficiency. For that purpuse, we focuson a recent model called WaveFlow [13], and measure theeffective energy consumption of training and inferring newsamples for five alternative configurations.

green computing

current re-searches focus on generation quality rather than on computa-tional performances when evaluating and comparing models.

The training time needed forthem to converge along with their complexity

aw waveforms requirehandling intricate temporal structures at both local and globalscales. Therefore, models are usually highly complex andcomputationally expensiv

eep generative models have reached an unprecedentedquality for waveform synthesis (in both speech and music).

which takes into account boththe quality of the model and its energy consumption

However,this race for quality comes at a tremendous computationalcost, which incurs vast energy consumption and greenhousegas emissions

iberation from mundane,menial tasks in these circumstances is tantamount to liberation from the ability tomake a living – and, more to the point, the ability to make a living as a musician.

Third, commercial applications using machine learning to generate cheap musicshould be a cause for concern, even if the only kind of music presently at risk is thehistorically stigmatised genre of production music.

As a consequence, what may seem like empty marketing hype at pre-sent may end up shaping the agenda for future work in this domain, encouragingcertain pursuits while suppressing others.

According to this line of argument,delegating to machines mundane and menial forms of creative work – like turningfinancial reports into news stories (Martin 2019) – might free up creative energies thatcould be more fruitfully directed elsewhere. Extended to production music, concedingthis domain to machines would presumably liberate musicians to pursue more aes-thetically rewarding activities

machine learning by certain of these firms hardly represents the most innovativeapplications of such technologies

this power would be redistributed to musicianswho, by and large, do not work directly for the companies in question, but whosemusic does.

One possi-bility would be the creation of some kind of ownership fund, either targeting individ-ual firms or the music technology sector more broadly. In line with other workerownership funds proposed over the years (Guinan 2019; Gowan 2019), shares mightbe issued to a body representing those musicians whose creative output is exploitednot just by music AI companies but other music tech firms as well. The main appealof such funds is that they redistribute not just wealth, but economic power, includingthe power to determine how and where to invest resources.

The proceeds would be directed to the Trust Fund, which then distributedthe monies raised to pay for free concerts across North America. Not only did this pro-vide underemployed musicians living outside major urban areas with paid work,redressing geographic disparities in cultural participation, but it also diminished some-what the winner-take-all tendencies that technologies of mass reproduction exacer-bate

The same principle holds for machine learning techniques, despite the distance sepa-rating them from the Markov processes employed by Olson and Belar. What ties themtogether is a reliance on what Adrian Mackenzie refers to as ‘probabilization’, as‘formalisms derived from statistics’

While certain trigrams are more probable and others less so, it’snot the case that an improbable sequence (like E4-D4-C#5) somehow counts for less,or that the single song where it appears contributes less than others. The song’s con-tribution isn’t the pattern, but its impact on the overall distribution of probabilities.

If it is difficult to isolate the contribution made by anysingle input, this is because no input contributes in isolation.

Again,within current copyright regimes this test applies only at the level of individual works.A prominent case in point is Robin Thicke and Pharrell Williams’ 2013 song ‘BlurredLines’. Following a lengthy lawsuit, in 2015 a jury found the two musicians guilty ofhaving infringed upon Marvin Gaye’s 1977 hit ‘Got to Give It Up’.

like the shared conven-tions governing a genre – to produce a technical resource

determining thecontribution their works made to its training, and apportioning royalties accordingly.Such difficulties would appear to rule out, either in principle or in practice, anyclaim that authors of training data might have on works generated by a machinelearner trained on their music.

Artist Rights Watch, for one, has called for musicians to invoke the marketingrestriction clause in recording and publishing contracts to refuse their music’s use ‘forAI purposes of any kind’

whose system istrained on a large number of musical ‘stems’ that an in-house composer in theiremploy creates for hire. Barring that, companies can assert some sort of exemption.

Crucially, developers of commercial systems, unlike academicresearchers, aren’t obliged to reveal the sources of their training data

Thatthese companies have title to the algorithms they developed isn’t in dispute; what is in dispute,however, is whether the works their systems produce belongs to them, some other party, ornobody at all.

A variety of legal doctrines have been mobi-lised in support of each of these candidates. Some have appealed to utilitarian theoryto buoy the claims of programmers and/or owners of AI systems, arguing that grant-ing them rights to AI-generated works will encourage the continued growth of the AIsector

Yet AIs, unlike humans, are insensible to such rewards,whether monetary or symbolic. Insofar as ‘machines need no incentive to work’,

Granting authors a temporary monopoly over their creations is regarded as animportant spur to creation, one that ideally harmonises individual and general interest:artists are rewarded for their investments of time, effort, and resources

What is more, there’s little appetite within legal circles for reforming statutesto grant machines rights on the works they produce.

in UK lawmachines aren’t eligible for copyright assignment, with rights defaulting to their creatorsor owners.

machine turned out, their recombination in different configurations

Another, less visible place wherethe same transition can be seen is in traditional production music companies, whichhave also adopted a platform model. But in contrast to these and other, more familiardigital platforms (like Facebook or Amazon), the platformization of commercial musicAI doesn’t involve one group of users being connected to another, but instead agroup of users being connected to an AI system.

he fact that onedoesn’t pay with money doesn’t mean one isn’t paying in some other way, usingsome other currency. As with so many other digital services, payment is still beingmade: it is simply that it is being made in the form of personal data

do not sell products to clients, but services. A case inpoint is Mubert, a company that bridges the consumer and business-to-business mar-kets. For brands, content producers, and/or brick-and-mortar businesses, Mubert offersa range of subscription plans. For a flat monthly fee, one can generate as muchbespoke music as one needs or desires ‘for free’

25,000 MIDI files on its site,in such genres as klezmer, tango, and the blues, while bitmidi.com boasts roughly113,000 MIDI files, from an equally diverse range of genres and styles.

Weav Run, whose appadjusts tracks according to the cadence of one’s stride whilst walking or running, withnot just the tempo of a track changing in real time, but also its texture, timbre, andarrangement (Weav Music 2019). A third example is AI Music, whose founder describesits applications as a means of ‘shape-shifting’ music so that it can adjust to differentlistening situations

suchmusic is not intended for direct consumption by end users, but is marketed instead toother cultural producers, typically for use in mixed media products like games, adver-tisements, or online web content.

the ‘MusicComposing Machine’ developed at RCA in the 1950s

one that transposes tracks from one musical genre to another

Since 2015 there has been a marked growth in the number of startups and technol-ogy companies seeking to commercialise music produced using artificial intelligence.

the article sketches a couple ofalternative models (levy-based trust funds, ownership funds) thatcould provide a more equitable institutional

the music that constitutes the trainingset necessary for machine learners to learn. Given the massivedatasets mobilised to train machine learners, existing copyrightregimes prove inadequate in the face of the questions of distribu-tive justice

recently dis-cussion has focused on who (or what) should be awarded rightsover the products of so-called ‘expressive AI’: Its programmers? Itsusers? Or the AI itself?