Festplatte mit Noten

More and more musicians are using artificial intelligence to create new pieces of music. AI can compose, write, arrange and generate sounds and voices. However, these models are trained using music that is, in many cases, copyright protected. The debate over the opportunities and risks of AI is currently in full swing.

Artificial intelligence, with all its promise, has now entered the creative scene. With just a click, a photo can be turned into a van Gogh, Obama can speak words that he has actually never said and AI can create entire pieces of music, with separate tracks for percussion, bass, guitar, keyboard and voice if we choose. Is there still a place in all this for human artists? Or will the latest compositions one day simply be generated at the push of a button by the AI systems of the Big Tech giants? In addition to worries such as these, the use of AI models in music is also a source of hope and has the possibility to create very real opportunities. These models are and have been used for quite some time by the music industry for production, performance and marketing as well as by independent musicians in their forays into new, experimental worlds of sound. AI can also be useful, for example in the area of singer songwriting, in reducing the initial hurdles for composing one’s own songs and it can serve as a tool for young musicians to complete management tasks. When it comes to the issue of accessibility, AI also has great potential to make it easier for people with disabilities to access music. 

Despite all of these positive aspects, it is important to take a critical look at how artistic content is handled that serves as the basis for generating new pieces of music. In this context, the term “machine learning” only describes the technical process by which an AI tool is trained using thousands of already-existing pieces, including information on aesthetics, form and the latest trends. Most of the compositions used for this type of training, however, are copyrighted, and, apart from the 2024 EU AI Act, which creates transparency requirements and calls for a summary of the copyrighted data used, there is currently no law that regulates this use of copyright-protected work. Court rulings are simply reacting to new trends, and the legislation is only adapted retrospectively to new possibilities and their abuse. But what type of legislation should even be passed – particularly regarding the global operations of the tech giants developing the AI tools?

Since the answer to this question will not only affect legal experts, awareness of the precise content of these laws needs to be raised in both public discussions and in policy-making. In Germany, various cultural institutions – including the German Cultural Council and the German Music Council – have called for legal provisions, which would regulate licensing agreements and compensation for the use of copyrighted content. In September 2024, GEMA (German Society for Musical Performance and Mechanical Reproduction Rights) was the first collecting society in the world to propose a licensing model for generative AI. The society also filed a suit against Open AI in November 2024 as well as against Suno AI in January 2025, in which it accused them of using copyrighted musical pieces to train their systems without acquiring a license. The problem remains that, until the appropriate legislation has been put in place, private companies are allowed to take advantage of the lack of regulations, putting the incomes of thousands of artists in jeopardy. If, as a result, they are no longer able or willing to create new training material for these machines to learn from it could have an effect on the continued development and innovation of music in general.

In order to first understand how a computer can create music, the following section will define and explain how AI works and describe its practical applications. Following this description, ethical aspects and the issue of copyright infringement will be discussed.

What is AI?

The field of computer science differentiates between strong and weak AI. If a human being were to communicate with a strong AI, they would not be able to tell whether it was an AI or a real human being (known as the Turing test). All of the systems that are known so far are what is known as “weak AI”. They show certain characteristics, mistakes or very specialized capabilities that quickly make them recognizable as a machine.

Intelligence, in general, “is the cognitive or mental ability of humans (...) to solve problems”. [1] These cognitive abilities comprise a number of specific qualities. Mathematic, abstracting, linguistic, logical, geometric, visual and auditory capabilities as well as factual knowledge are often necessary in order to solve a specific problem. In contrast to human intelligence, AI systems are highly specialized. They can solve some problems incredibly well, while completely failing at others.

Unlike human intelligence, artificial intelligence is carried out by machines and software. It attempts to approximate human capabilities as closely as possible and simulates strategies of the human brain. Thus, AI is an “automation of intelligent behaviour that makes use of machine learning”. [2] This automation is achieved via neural networks, which have been set up using programming language and which scan the data entered looking for patterns.

How AI systems work

AI systems get their creative capabilities largely through the recombination of learned facts that they have gathered during the machine learning process. This data is then evaluated using patterns, probabilities and frequencies. Patterns are created through the accumulation of combinations of various characteristics. Specific data combinations will appear in a certain order more often than others. But how does this data analysis result in a behaviour that can be called “creative” or “intelligent”?

AI from the 18th century: a musical dice game
One way to understand how an AI can be creative is by looking at the musical dice game often attributed to Wolfgang Amadeus Mozart. In this game, precomposed patterns are created (bars of music) that are suitable for certain harmonic sequences: the beginning, middle and end of the piece. For each bar, there are a specific number of variants. Using a die, various versions of the musical bars are selected and combined. The compositions created by the game are all different and all sound coherent and harmonious, although they were created by pure chance. Here, chance produces the creativity by putting together a new combination each time. The pre-prepared bars and their position in a 16-bar waltz contain the order and the existing pattern. This order and the combinations chosen by chance mean that, by using this process, the 176 bars prepared by Mozart have the possibility to create approximately 45 million different versions. [3]

Title page of: Anleitung so viel Walzer man will mit Würfeln zu componieren ohne musikalisch zu seyn oder Composition zu wissen, Berlin [ca. 1790]
Title page of: Anleitung so viel Walzer man will mit Würfeln zu componieren ohne musikalisch zu seyn oder Composition zu wissen, Berlin [ca. 1790]  
Excerpt from Mozart's Musikalisches Würfelspiel (musical dice game), from: Anleitung so viel Walzer man will mit Würfeln zu componieren ohne musikalische zu seyn oder Composition zu wissen, Berlin [ca. 1790]
Excerpt from Mozart's Musikalisches Würfelspiel (musical dice game), from: Anleitung so viel Walzer man will mit Würfeln zu componieren ohne musikalische zu seyn oder Composition zu wissen, Berlin [ca. 1790]  
Table of numbers for the waltz, from: Anleitung so viel Walzer man will mit Würfeln zu componieren ohne musikalische zu seyn oder Composition zu wissen, Berlin [ca. 1790]
Table of numbers for the waltz, from: Anleitung so viel Walzer man will mit Würfeln zu componieren ohne musikalische zu seyn oder Composition zu wissen, Berlin [ca. 1790]  

A simple AI: Markov chains
A more complex example can also be used to explain more precisely how neural networks function. Markov chains were developed at the beginning of the 20th century by the mathematician Andrei A. Markov to generate variants with different degrees of similarity to the original. This system can be used for visual as well as musical patterns and can be considered a simple form of AI, as is shown in the following example.

The song “Happy Birthday” contains the following notes:

FIGURE 1
Notes of "Happy Birthday"

If we analyse these notes according to their frequency, measured by percentage, we would get a general description of the melody. As a first-order Markov chain, it would look like this:

FIGURE 2
First-order Markov chain of "Happy Birthday"

In the next step, this analysis would be used to generate a new melody. Weighted chance is used to do so, and, although the selection is random, c4 will make up 32 per cent of the resulting notes and d4, twelve per cent. The result is a tone sequence that is similar to the original but sounds rather random. In the next step, the analysis is refined. Now, the frequency of two consecutive tones is examined. In the example below that would mean that there is a 37.5% chance that the note following a c4 is also a c4 and a 25% chance that it is a d4. When this analysis is used to create a new tone sequence, the result is certainly recognizable as “Happy Birthday” even if it differs considerably from the original. The more 100-per-cent values the analysis yields, the more similarities exist between the result and the original melody. In the following analysis, there are already three 100%, one 66.7% and four 50% probabilities of tone sequences.

The second-order Markov chain then looks like this:

FIGURE 3
Second-order Markov chain of "Happy Birthday"

Various analyses can now be carried out that each examine the sequence of two, three or more tones. The following shows an analysis with the probability of four consecutive tones. The high percentages shown make it immediately evident that the melody generated by this analysis is a copy of the original and only contains 100-per-cent tone sequences.

Fourth-order Markov chain:

FIGURE 4
Fourth-order Markov chain of "Happy Birthday"

In conclusion, the findings indicate that when the data is generated using AI rules, characteristics and patterns in the data are analysed and configured in a new manner using a varying degree of random chance. A neural network, however, analyses a much more complex matrix of characteristics in order to generate the requested data. In the procedure described above, only one parameter – the pitch – has been analysed and weighted using a probability expressed as a percentage. A neural network, on the other hand, would also analyse patterns in the volume, chords, harmonies, accompaniments, tone quality, articulation, etc.  This example provides an idea of the enormous amount of data that results from the neural networks’ analyses and just how complex the network of these extremely varied patterns is in this type of analysis. The goal is not to obtain a copy of the original information but rather to identify the most important characteristics in a piece of music and use them to generate new information. In this type of analysis, the specific characteristics of various melodies, information and pieces of music are mixed together. The result usually appears to be novel although it is nothing more than a randomly selected mix of previously-existing musical patterns.  

The history of AI in music

Research into artificial intelligence (AI) began back in the 1960s in the context of computers and programming language. The mathematical neuron mode, which is fundamental for AI, had already been presented among experts back in 1943. However, AI tools with neural networks were not well known in the general public until the 2010s. One of the first types of AI tools to be used in music was the Magenta Project developed by Google. It was launched in 2016 and its software has been open source since 2019.

After the Magenta Project was launched, things seemed to be quiet for quite some time regarding AI and music. The reason for this silence was the intensive research and development work that was invested in the successive step. There were many additional components that needed to be researched in order for AI to process and generate sound. Currently (as of 2024), AI is capable of communicating with humans via language and can even make a human voice sing convincingly.  It is quite natural then that music was one of the first fields to be used in creativity research and AI. After all, composers use rule systems and algorithms in their pieces when they themselves are creating music. Throughout the time that computers have existed, the ways to use rule systems and the complexity of these systems have only increased. Thus, it is not surprising that Markov chains were used back in the 1960s by composer lannis Xenakis as a simple AI in the composition of instrumental pieces or that, in the 1980s, functional AI systems could compose pieces emulating the styles of Vivaldi, Bach and Mozart. Back in 1983, American composer David Cope was able to compose Bach-style chorales without neural networks using his EMI (Experiments in Musical Intelligence) systems. Later versions of EMI were able to independently analyse music in order to compose new pieces. They anticipated the machine learning of the neural networks to come. In addition, various chance-based, stochastic processes, such as Markov chains, were used, in which rules and chance were mixed.

AI with neural networks

Infrastructure, big data, digitisation in society: prerequisites for the boom
There are various components responsible for the massive presence of AI over the last 15 years. On the one hand, digitisation is increasing in scale and speed. Digital devices and, in particular, storage technology for data have become cheaper and faster. Another important aspect is the development of network technologies that can transfer large amounts of data via fibre optic cables, WIFI and mobile networks. The mass-scale use of devices such as smartphones, cameras and laptops as well as services such as Google, YouTube, etc. creates large quantities of data that have to be stored in an ever-growing number of processing centres. The world’s biggest digital firms such as Microsoft and Amazon have turned the sale of services combined with data storage into one of their most important business areas.

The increase in data, however, necessitates better data processing strategies in order to use it to its full potential. Neural networks offer an exceptionally powerful option to make the content of these large quantities of data usable: data mining. The possible applications for these networks have been decisive in how widely they have been used. As long as they were only used in processing centres, the only people who could interact with AI systems were those with significant computer expertise. In 2022, however, ChatGPT made a breakthrough in language-based methods which use Large Language Models (LLMs) to formulate written and verbal requests (prompts). This language-based interface between humans and machines significantly reduced the barriers to accessibility when it came to AI interaction and led to its continued spread and development.

How humans talk with machines
After ChatGPT, many other powerful AI systems came onto the market that are now capable of composing and arranging music according to a prompt. One of the most important components for the success of these technology-based systems is the way that humans can communicate with machines. This aspect of Human-Computer Interaction (HCI) has been revolutionised by AI systems in a spectacular manner. The interaction with the machine through free, creative language is one of the greatest achievements of the last AI wave. It means that every person who speaks the necessary language is able to operate this machine. Prompting can also be done in written form. A prompt gives the machine a task to complete. If the machine is to carry out complex tasks, it is possible that users will need to be trained in how to prompt it appropriately. Thus, the profession of prompt engineering has been created, in which strategies are developed in order to operate AI even more precisely.

How neural networks work
In the past, algorithmic processes were used to analyse the content of databases to find specific patterns. Neural networks, on the other hand, make it possible to analyse previously-unknown quantities of parameters according to incredibly complex patterns. Neural networks are capable of analysing data in an exceptionally flexible manner and can independently combine many characteristics of the data with one another. Without knowing which pattern it should be looking for, a neural network can find characteristic patterns in the data it has been provided to analyse. It is this analysis alone that makes this data valuable for its owners or for use in an AI.

In surveillance technologies – an area in which AI has been used for quite some time – it is important to identify conspicuous or abnormal patterns in movement. In the past, it was necessary to have people with special training who would watch numerous cameras looking for these conspicuous patterns. A neural network was then first trained to recognise objects and then conspicuous patterns in their movement. This was then later combined with neural networks for facial recognition, which made it possible to identify people and their movements in a specific area. In order to handle such a complex task, various capabilities are necessary and, oftentimes, AI systems that are connected with one another. 

This division of labour is precisely how the human brain functions. There are specific regions that are responsible for processing and optimising signals from acoustic, optical and olfactory sensors. The process of recognising, differentiating and identifying the objects and then comparing them with data that is already saved in our memory, with which a recognised object can be classified and remembered, is carried out in another region of the brain. We hear a noise – the vibrations are recognised in the brain stem as disassembled frequencies and are then analysed according to acoustics, direction and other characteristics. Then, the patterns that have been recognised are separated out, since various other objects could also be in the room, and they are compared with the known patterns.

Similar to the way our brains work, ChatGPT also use this type of division of labour. The process, put simply, looks like this: if the AI receives a command in spoken language, a speech recognition AI is used to translate the spoken language to letters. The command is analysed and is sent to the actual AI with the archived content. This AI will carry out the task and then the answer will be composed in language and forwarded to a language synthesis AI in order to for the answer sound like a voice.

Many processes in neural networks are so complex that we currently only have an elementary understanding of them. Whether the AI can provide the desired results and whether the black box of the neural network is working properly can only be determined by the input prompt and the quality of the answer. This is one of the reasons why it is difficult to undo mistakes within an AI system. If the data is skewed or incorrect, it is not currently possible to correct it to the desired extent. 

Fields of application

Voice Cloning
In theory, the areas in which AI systems can be used are practically limitless. However, in reality, there is a lot of development and optimization work that needs to be carried out before a system can work well. It is not always clear whether an AI system will be more efficient and reliable than a system based on the current standard technology. What makes AI systems really attractive, though, is when they can carry out tasks that were previously not possible or could only be done less efficiently.

One example of this is voice cloning, which is when voices or instruments are extracted from a recording, as was done for the Beatles song released in 2024 “Now and Then”. To create this song, the voice of John Lennon, who died in 1980, was taken from an old recording using AI and then mixed with other pre-produced elements. In this method, any number of instruments as well as all recognizable noises can be removed from a complex sound recording, which is known as “denoising”. 

Video

A lot of research has been done in the area of language generation. Currently, a speech AI is now capable of using relatively little learning data to make a person’s voice speak any kind of sentence and even sing. The use of this simulation, however, is problematic due to personal rights when real people are involved. [4] There are some actors as well as singers, such as Holly Herndon, who have given their voices to the public, so that they can be used to experiment.  On the other hand, though, some Hollywood actors have gone on strike over the issue, calling for the unauthorized use of their artificial figures and voices to be regulated. This area of dispute includes not only the reproduction of language and voices but also the picture and movement data of people and how this data is saved during the process of machine learning and reorganised by the AI.

Music sound optimisation
AI systems really show their strength whenever specific patterns need to be reproduced or recombined in a new manner. For example, to master previously existing pieces, clearly defined sound patterns can be generalised and applied to other already existing pieces. The piece is processed by the AI and, with the help of some additional fine tuning, the sound of the piece is then optimised. The sound patterns of pop music, which are optimised for headphones and mobile phones, are comparatively simple. Experimental electronic music, on the other hand, which uses unpredictable, novel sounds, is more complicated. This is where AI reaches its limits, since the patterns do not work with granular synthesis or noise sound music. It is also possible that the personal touch musicians lend their own pieces will be lacking when mastering software is used. The standards should not be too strictly formulated, otherwise listeners may not like the music or find that it sounds too stereotypical.

AI samplers
AI-based samplers could certainly provide some advantages when it comes to simulating real instruments. In terms of tonal quality, samplers today have certainly come a long way, but they do have a few weak spots. Most of the time, only individual tones of an instrument are recorded, which means that the interpolation between the tones is missing in a melody with multiple consecutive notes. These generally subtle and quiet interpolations between the stable vibration states of the instrument can be heard in the bowed, plucked and wind instruments. However, this effect can also be clearly heard when tones are repeated played by a piano. If a string of the piano is already vibrating and it is hit again, slight changes in the timbre of the pulse wave are perceptible. These interpolations are not reproduced by samplers. Only physical instruments can create these types of micro-interpolations. However, they have to be optimised for each individual instrument, which is a time-consuming process. An AI system could also generate these characteristics and possibly make use of individual instruments played by the user. In addition, instrumental playing styles, e.g. a human filter, could be integrated into the generation process in order to reproduce the physical possibilities of playability. There are numerous possibilities in the area of AI-based samplers: developing virtual instruments, combining different instruments, projecting the playing style of a harp to a bowed instrument, etc.

Cloning
The possibilities to use characteristic properties are not only limited to instruments; they can also be used for performers. [5] Cloning is already being used for avatar artists that have been completely artificially created. To do so, real movement data is used in order to ensure that the avatar’s movements appear natural. The degree of virtualisation of artists is particularly advanced in Asia. There, there are already figures that function as realistic avatars. In essence, this is a way to fully clone a person using AI and virtualisation technologies to create a clone of a person’s appearance, voice and music. [6]

Pianist Francesco Tristano at the AI project Dear Glenn at Ars Electronica 2019
Francesco Tristano at the AI project Dear Glenn at Ars Electronica 2019  
Photo:  vog.photo
Scientist Akira Maezawa at the AI project Dear Glenn at Ars Electronica 2019
Scientist Akira Maezawa at the AI project Dear Glenn at Ars Electronica 2019  
Photo:  vog.photo
Flutist Norbert Trawöger and violinist Maria Elisabeth Köstler at the AI project Dear Glenn at Ars Electronica 2019
Flutist Norbert Trawöger and violinist Maria Elisabeth Köstler at the AI project Dear Glenn at Ars Electronica 2019  
Photo:  vog.photo

Prediction of musical preferences
Recommendations from AI systems that predict our musical preferences are already a part of our daily lives. The recommendation lists from streaming services are nothing new. However, recommendations can only be given when the system can recognize the characteristics of a piece of music and can also determine which characteristics the listener prefers. To do so, the music must be automatically assigned keywords such as “calm”, “fast”, “classic”, etc. – known as Automatic Information Retrieval. Only then can the system determine where there is an overlap between the musical characteristics of a song and the predicted musical taste of the user. When a database contains millions of songs, it makes sense that this type of work is best carried out by a complex AI system.

Composition platforms
For a long time, algorithmic composition using mathematical strategies and stochastics as composition tools was widely used, but now the capabilities of composition platforms have radically changed. Since the end of 2023, a new type of creative AI has been available that not only makes compositional decisions but can also create sound and entire arrangements. These complex systems include platforms such as Udio, Suno and Stable Audio as well as others that use prompts to produce compositions with durations of a few minutes. They are also capable of supplementing sound files that users have created and uploaded and to augment the pieces using the tonal characteristics of the uploaded samples. Even when the users’ stipulations are vague and imprecise, the AI in the platform is able to supplement everything that is missing using information with the highest probability, thus producing complete musical pieces that are of excellent quality and have great stylistic variation. 

Composition AI can provide ideas and examples even if, for a variety of reasons, it is not yet usable for most professional purposes. For example, the timing control is currently insufficient for film composers, the tonal details are too difficult to control for sound logos and it is completely unsuitable for large formats. Platforms such as Udio and Suno generate results that are practically unusable for composers, since the components themselves cannot be processed. For their own work, composers and arrangers prefer to use AIs that generate MIDI files. It can however be assumed that all of these tools will continue to be developed and, as long as there is a market for it, will be adapted to the needs of their users. Especially in popular music, there are currently many musicians who integrate AI vocal lines prototypically or even directly into their compositions. In the charts, there are already international and national hits that have been almost completely created using AI tools.

AI in the classroom
Sound analysis can also be used in the classroom or in musicology, e.g. to analyse written compositions and find mistakes or specific characteristics, which provides the students with immediate feedback. AI can also be used to inspire students’ creativity, regardless of their level. This area of application depends on the competency of the school systems and teachers. It also depends on the way they choose to integrate AI-simulated creativity into lessons in order to encourage students to create music, improve their listening skills and to overcome their own hurdles to musical creativity.

Musical creativity for all
AI-based music platforms lower the demands on users Everyone can verbally express musical characteristics and thus also create music. The music created in this manner can indeed reproduce existing styles and can mix characteristics to create songs. However, an AI is not really capable of producing substantially new music or new trends. These platforms will therefore not be able to replace composers, but they can be developed into powerful tools in order to speed up production processes. Creative artists have made and will continue to make all of the content that AI systems rearrange. The question remains, though, of how the public will react to music that has been produced with these new tools. Through prompting and uploading human-made sounds, people can control the AI’s reproductive creativity and have the possibility to steer it in innovative directions. In the more conservative music scenes, a bit of innovative inspiration from an AI system will be a welcomed tool; in more experimental areas, it will reach its limits and leave us wondering what exactly creativity is.

„An AI will never be aware of the emotionality or the aesthetics of music, since, unlike humans, it does not possess intention or feelings that allow it to perceive music."
Autor
Ludger Brümmer

Human vs. machine creativity

During their studies, composition students analyse works in order to identify their rules and characteristics. They then use these rules to create their own compositions. This practice is actually comparable with the process used by an AI system based on a neural network. During the process of machine learning, an AI analyses numerous works in order to obtain a large number of patterns that it can use. After it has been prompted, it then chooses some appropriate examples from these patterns, combining them in a new manner to create a new composition.

What is interesting here is the question of artistic intention, which is the starting point for the creative process. How can something new be created from the compositions that have been studied which is more than just an intelligent rearrangement of old patterns? Humans also weigh the importance of information according to their own needs in the form of sentiment, emotionality or the desire for novelty. They engage in the creation of a new work through the experience of its construction, through surprises or calculated planning – in short: through the interaction with the creative process.

AI systems (still) differ from the human creative process when it comes to the quantities of samples that are necessary for machine learning. Currently, AI systems have to evaluate hundreds of thousands or even millions of compositions in order to generate high-quality output. For humans, on the other hand, a few hundred will suffice. They can prioritise works or specific information much more when it comes to the expected meaning or individual taste and can thus optimize the work process. Other aspects, such as recombining information and patterns, can be found in human as well as machine creativity. 

However, an AI will never be aware of the emotionality or the aesthetics of music, since, unlike humans, it does not possess intention or feelings that allow it to perceive music. They can only gauge how frequently specific patterns appear in an immense amount of data and gather information on which characteristics these patterns have.

Ethical questions of machine learning

Machine learning uses content, analysing the information it contains and weighting its constellations and probabilities. In other words: AI generates a statistical description of the material. This description is then saved with thousands of other descriptions and serves as a template to create new compositions, which is then mixed with so many other similar characteristics of other pieces that it appears as if something new and novel is produced from this differentiated mix. The AI can be specifically commanded to do so by using the input prompt “Create a composition in XY style”. 

The problematic aspects of this new way of using copyrighted information require us to think about the process in a new way. Copyrighted material is being saved and analysed. This information is being retrieved by users and can either result in a copy of the piece or a complex mix containing details of many different compositions. The use of composition patterns – whether it is tonal, rhythmic or melodic patterns – is, however, subject to authorisation and/or remuneration according to German law. [7]

On the other side of the spectrum are the pieces that belong to the cultural public domain. Genres as well as other additional information and patterns are not copyrighted, nor are pieces written by composers who have died more than 70 years ago. 

What is it exactly that differentiates these creations from those of composers who take inspiration from Stockhausen or Boulez in creating their own pieces or take standard jazz pieces and use them to make new compositions? In any case, the use of random chance and patterns is always an aspect to be considered when it comes to creativity.

It becomes increasingly difficult to determine the basic difference between human and machine, since terms such as “consciousness” and “intelligence” have not been precisely defined in the sciences. We ourselves do not exactly know what our consciousness is, which makes it difficult to say at which point a machine also has one. 

However, there are some fundamental differences: composers would make decisions with intention, combine them with emotions and would be capable of creating something new that would be completely the opposite of everything they had previously created. The leap from baroque to classic, from there to the romantic and finally from tonality to atonality would not be possible with today’s AI systems, since they are only based on stored identities and information. Leaps such as the one from tonality to atonality are true paradigm shifts and thus an expression of creativity that is not rooted in functional statistics and mixing characteristics but rather in individuality and dialectics. This type of creativity is something that we simply cannot currently expect from an AI.

MIZ KNOWLEDGE

Further sources

Dokumente

Title
German Cultural Council calls for clarification of copyright issues in the use of AI (01/2025)

Text
The German Cultural Council’s position paper on copyright issues in connection with AI. The council urges the responsible parties to act quickly in the interest of the creative artists affected. (in German)

Title
Initiatives for ethical development and use of AI in cultural and creative industries (11/2024)

Text
In collaboration with international experts, the German Commission for UNESCO has developed initiatives that are designed to ensure that the use and development of AI in art, culture and the creative industries is carried out according to ethical principles and that existing international law is adhered to.

Title
GEMA’s AI charter (11/2024)

Text
In its AI charter, GEMA defines ten ethical and legal principles for fair and sustainable interaction of human creativity and generative artificial intelligence.

Title
Artificial intelligence and music – a useful tool or competition? (10/2024)

Text
In light of the upcoming creation of the contents for the EU AI Act, in 2023, the German Music Council issued a list of demands to the German Federal Government and the federal states, which included the areas of copyright protection, personal rights and digital education. In October 2024, an update to the living paper was published. (in German)

Title
Artificial intelligence: sustainably developing parameters for art and culture (10/2024)

Text
The German Cultural Council’s position paper on the use of artificial intelligence in the art, culture and media sectors, which addresses issues of how AI should be handled and regulated. (in German)

Title
Joint position paper from DTV and VERSO (06/2024)

Text
The German Lyricists Alliance (DTV) and VERSO, the German songwriters association, demand transparent payment models for streaming platforms as well as clear regulations for the use of copyrighted music data to train AI tools. (in German)

Title
AI and Music (01/2024)

Text
In the world’s first study on the topic, GEMA and its French affiliate SACEM investigate together the effects of generative AI on the music and creative industries.

 

Title
AI ACT: open letter to the federal government written by Germany’s cultural, creative and media industries (01/2024)

Text
A broad coalition of associations appeal to the federal government to approve the draft of the EU AI Act. (in German)

Title
Artificial intelligence calls for maximum transparency (09/2023)

Text
Within the context of a legislative proposal by the European Union to regulate artificial intelligence (AI Act), the Cultural Council NRW has written a position paper in which it lists the conditions for artists that must be considered. (in German)

Title
Formulation suggestions for the AI Act (version: 19 September 2023) (09/2023)

Text
In this position paper, the Author’s Rights Initiative calls for clarification on regulations in the European Parliament’s AI Act.

Title
Artificial intelligence and copyright – position of the German Cultural Council (06/2023)

Text
The German Cultural Council asks the German federal government to look into whether the current statutory exemptions for text and data mining also apply to the use of copyrighted materials to train AI systems. (in German)

Title
AI from the perspective of music authors and performing artists (Composers Club, DEFKOM, DKV, DTV, mediamusic, Pro Musik & unisono in April 2023) (04/2023)

Text
In this position paper, music associations such as Pro Musik and unisono make suggestions for regulating AI to protect copyrighted musical material. (in German)

Title
Authors and artists call for measures to protect against generative AI in the European AI Act (04/2023)

Text
In an open letter, the Author’s Rights Initiative calls for a stop to the improper use of copyrighted material in the creation of AI-generated data and material.

Title
Tutorial: basic questions on copyright in music

Text
The miz tutorial “Copyright in music” examines fundamental questions surrounding this topic and also delves into issues arising due to artificial intelligence. (in German)

Outlook

We cannot currently predict the limits of technological AI. Composition systems are currently capable of creating entire compositions and virtual people, albeit under questionable copyright conditions. If we take things a step further, AI systems are perhaps even superior to humans. Through the possibility to separate content and data from the hardware, an AI can, in theory, “live” forever. If the hardware breaks, it can be replaced and the content of the system can be copied to the new hardware. In addition, AI systems are also capable of learning in one location and passing on the information they have learned to other systems, and they can do so at a speed that is far faster than any human could teach or learn. 

AI systems can also process data through recursive methods whose potential to optimise quality is far beyond the capabilities of a human brain. The systems do not yet have sensors that would allow them to keep track of societal processes, although they are continuously updated on current developments via the internet. What would happen if AI systems were able to learn new musical developments and could use them to make suggestions and prognoses for new styles? What if an AI was connected to radio-station playlists and streaming platforms as well as being fed material from current musical pieces? Would this type of AI know what music was trending and would it be able to independently make prognoses?

The debate surrounding copyright issues and the legal regulations for the use of copyrighted material to train AI systems have only just begun. With the AI Act, the EU has already implemented new legislation on this issue and it is evident that views on the use of copyrighted material are beginning to change. In the U.S., some AI platforms were sued by major labels in 2024 for their use of copyrighted material. In the same year in Germany, the GEMA was the first collecting society in the world to file a test case in this area. But what would happen if the operators of these platforms were to simply move to areas with political systems where copyright protection does not exist? One of the few platforms that has been transparent on this issue is Stable Audio, which has made public the sources of its training data. According to this publication, only data was used for which users had provided a consensus declaration. Perhaps this can be seen as a model for the future. 

Important voices in this discussion include the interest groups for musicians and composers, which include the GEMA, the German Cultural Music Council and the German Music Cultural Council but also the cultural, creative and media industries which stipulate the essential need for comprehensive remuneration for every use of copyrighted material and which use events as well as position papers to call attention to this fact in political and societal discussions of the matter. This also includes suggestions for feasible and practical billing systems with micro-payments, in which every composition used would receive a payment, albeit a small one for a limited period of time, such as is already the case for YouTube and streaming platforms. 

Currently, AI systems are tools that can be used to create music. They develop suggestions, hypotheses and serve as an interesting place of experimentation for musicians. They help implement ideas and can also be very useful for tasks that are particularly standardised. However, the question arises of whether musicians will still need knowledge in the subjects such as counterpoint or instrumentation, since AI can easily take over these tasks. Currently, AI-generated music has not been as magical as some may have predicted.  

It may put us at ease that AI systems are not yet capable of creating real innovation or developing aesthetic hypotheses. For example, it would have been impossible for an AI to make the leap from tonality to atonality, as happened at the beginning of the 20th century. This was a hypothesis that was qualitatively new, but, from a dialectical perspective, was simply a logical reaction to the preceding development of tonality. On other words: the consequence of a process with a qualitative leap as a result of the reflection. Current AI systems are simply ill-equipped to make these types of decisions. But perhaps someday, with the right prompt, it will actually be possible for an AI to implement such a paradigm shift.  

Über den Autor

Ludger Brümmer is a professor of composition with digital media at the Trossingen State University of Music and, for the last 20 years, has headed first the Institute for Music and Acoustics followed by the Hertzlab at the ZKM (Center for Art and Media Karlsruhe). As a composer and researcher, he works in the fields of physical sound modelling, algorithmic composition, spatial sound, artificial intelligence and interdisciplinary art. Ludger Brümmer has received numerous awards, including the Golden Nica at the Ars Electronica for Digital Musics & Sound Art, two Pierre d'Or at the Bourges Synthesis Competition and the Busoni Prize of the Berlin Academy of the Arts.

Footnotes

  1. cf. Wikipedia: Intelligenz. (in German) Online at: https://de.wikipedia.org/wiki/Intelligenz (accessed: 14 January 2025). 

  2. cf. ibid.

  3. cf. Musikalisches Würfelspiel – Kirnberger/Mozart. Online at (in German): https://www.youtube.com/watch?v=fK2MCXpDWB4, Mozart – Musikalisches Würfelspiel – 3 Ergebnisse. Online at: https://www.youtube.com/watch?v=u0Tin1s_GZk or Mozart Musikalisches Würfelspiel – 365 gewürfelte Walzer – 365 rolled Waltzes. Online at: https://www.youtube.com/watch?v=AP4BndvunKE (accessed: 14 January 2025).

  4. Cf. the dispute between OpenAI and Scarlett Johansson: OpenAI verordnet KI-Stimme eine Pause. (in German) Online at: https://www.tagesschau.de/wirtschaft/digitales/openai-scarlett-johansson-stimme-100.html (accessed: 7 January 2025).

  5. See, for example, a cloning project of Glenn Gould. Yamaha: Dear Glenn. Glenn Gould as A. I. Online at: https://www.yamaha.com/en/stories/new-values/dear-glenn/ (accessed: 14 January 2025).

  6. The U.S. broadcaster Channel 1 has cloned the moderators since 2024.

  7. For more information on this topic, see the German Music Information Centre: Tutorial Urheberrecht in der Musik. [Tutorial of copyright in music] (in German) Musik verwenden, bearbeiten und aufnehmen (Accessed: 14 January 2025).