The three big debates around music and generative AI
Over the past few years, there have been many discussions about music and AI, especially generative AI models that create music. These models are trained using existing recordings and songs made by people. They raise many questions for music creators and the music industry. Some are about the law, some are about money and business, and others are about what is right and wrong.
On some of these questions, much of the music industry agrees. Artists, musicians, songwriters, composers, studio producers, organisations that collect music royalties, major record labels, independent record labels and music publishers are mostly on the same side. But on other questions, different groups in the music community increasingly disagree.
With so many questions and issues, it can be hard to understand the debates about music and AI. To make things clearer, CMU has grouped many of these issues into three main debates about music and AI.
This Friday, CMU’s Chris Cooke will join a panel to discuss music and AI at Un-Convention in Manchester. Before that discussion, he explains each of the three main debates, including the key questions people are asking and recent changes and developments.
Big debate #01: What AI companies must do
On the input
To train a generative AI model to create music, an AI company needs to copy large amounts of existing music, including songs and recordings, into a training dataset. The company then studies this data to build its model. During the training process, it may also make and change additional copies.
Copyright law gives copyright owners control over who can copy their work. In theory, this means an AI company must get permission from all the relevant copyright owners before making these copies. For a recording of a song, this includes the owners of the separate copyrights in the recording and in the song itself, including its music and lyrics.
However, many AI companies trained their generative AI models without first getting permission from the rights holders. Instead, they argued that they could copy existing content, including music, without permission because of legal rules called ‘copyright exceptions’ or ‘fair use’.
Some countries have a copyright exception for text and data mining, often called a TDM exception. This allows third parties to copy content for this purpose without asking rights holders for permission. Many AI companies argue that this rule also covers AI training. They say that if they train their models in these countries, they do not need permission from rights holders.
US copyright law has a similar idea called ‘fair use’. If the use of a copyrighted work is considered ‘fair’, permission is not needed. Many AI companies argue that training AI models counts as fair use.
People who lobby on behalf of many AI companies have also been putting pressure on lawmakers in different countries to create new copyright exceptions. These changes would make it easier to use existing works to train AI without getting permission from the copyright owners or creators.
Most of the music industry agrees that these efforts should be stopped. The industry, along with other creative industries, strongly argues that existing copyright exceptions and fair use rules do not cover AI training and should not be changed to cover it.
To support this position, record companies, music publishers, royalty-collecting organisations and individual music creators have filed lawsuits in the US, Europe and other places. They want courts in different countries and legal systems to set precedents showing that AI companies cannot rely on copyright exceptions or fair use when copying music into training datasets to build generative AI models.
AI companies facing lawsuits brought by the music industry include Suno, Udio, Anthropic and Open AI. All these cases are still going through the legal process.
There have also been efforts to clarify or change copyright law. The aim is to make sure that even if AI companies can use exceptions in a small number of countries to train models without permission, they cannot then sell or offer those models commercially in other countries where those exceptions do not exist.
On the output
Some generative AI platforms also use existing music when creating their output. For example, they may remix existing songs. They may also use an artist’s voice or name by creating vocals that sound like a particular performer or letting users generate music ‘in the style’ of specific human creators.
When platforms are designed to do this, although it can sometimes happen by mistake, most AI companies accept that they need permission from the relevant rights holders and/or performers.
They generally agree that copyright exceptions and fair use do not apply when AI output copies or remixes existing songs. If AI output imitates an artist’s voice or creates music ‘in the style of’ an existing artist, it may also use other rights belonging to creators. These can include rights over a person’s identity or public image, rights relating to personal data, and trademark rights.
It is still not clear exactly how rights over a person’s identity, personal data and trademarks apply when AI output uses someone’s voice or name. In the UK, there is currently no general legal right over a person’s identity of this kind. Because of this, different countries have seen proposals for new rights covering digital copies of people, specifically when these are used in AI-generated output. The music industry has supported most of these proposals.
However, even as these rights are being clarified and developed, most AI companies are more careful about using existing works, voices and names in the content their models create.
Trends + developments
The main legal dispute between the music industry and the technology sector over what AI companies must do is far from settled. However, some technology companies have started negotiating licences with music industry rights holders. These licences cover both AI training and the music or other content created by AI.
There are several reasons for this. Some AI companies want to do the right thing. Others are also music streaming services and already have relationships with the music industry. Some want to offer tools for remixing music or copying voices, so they recognise that they need licences. Others are acting because of legal risks.
Whether AI training counts as ‘fair use’ under US copyright law, allowing existing music to be used for training without permission, remains an active and highly disputed legal question.
Dozens of lawsuits in US courts focus on fair use. These include cases brought by authors, book publishers, newspaper owners, film studios and the music industry. In the first two cases to reach trial, involving Anthropic and Meta, both of which concerned book authors and written works, the judges accepted the technology companies’ fair use arguments.
However, in the Anthropic case, the judge said AI training counted as fair use only if the company had obtained the original copies of the books it used from legitimate sources. Many AI companies have relied on pirated material to build their training datasets.
The judge in the Meta case also said an AI company might not be able to defend itself by claiming fair use if a copyright owner could show that the output of a generative AI model weakened the market for their work and reduced its commercial value.
For music, this means that if AI-generated tracks could be shown to significantly reduce the commercial value of the human-created music used to train a model, the company might not be able to rely on fair use.
These different court decisions have left AI companies unsure about the law. Under US law, a copyright owner can sue for up to $150,000 for each work that was deliberately infringed. Most AI companies have copied millions of works without licences. This means that if an AI company ultimately loses a copyright lawsuit in the US, it could face damages of a trillion dollars or more.
Given this risk, an AI company could argue that it is in its own interests to arrange copyright licences now. Some rights holders, including music companies, have used the risk of extremely large damages to put pressure on AI companies to negotiate licensing deals.
This leads to the second big debate.
Big debate #02: What record labels and publishers must do
On the input
Making AI companies agree to licences before they train their models sounds like a big win for the music industry, and in many ways it is. But as record labels and music publishers begin signing deals that let AI companies train models using their music catalogues, new questions arise for people and businesses across the music industry.
- Will labels and publishers ask individual artists and songwriters for specific permission before allowing their recordings and songs to be used in AI training deals? Older recording and publishing contracts were written before these uses of music were expected.
- How will labels and publishers divide the money they receive from AI companies among the recordings and songs in their catalogues? One argument is that every work used in AI training is equally important and should receive an equal share. But is that how the industry plans to divide the money?
- Will recordings and songs receive equal shares? With streaming, the recording usually earns more money than the song itself. With music used alongside film, TV or other visual media, known as synchronisation or sync, the two separate rights usually earn roughly the same amount. Will AI training income be split equally, or will another method be used?
- Once a label or publisher has assigned money to a particular recording or song, how much of that money will go to the artist or songwriter? Record contracts often set one royalty rate for streams and record sales and a higher rate for licences. Will the higher rate apply to AI deals?
- Will session musicians who played on a recording receive a share of this new AI income? They usually do not receive royalties from streaming or record sales.
Some labels and publishers have started contacting artists and songwriters, or their managers or lawyers, about these issues. But there is still very little public information about how the deals work, and no standard approach that everyone has agreed to follow. The AI deals themselves are kept secret under non-disclosure agreements, or NDAs.
The major labels and publishers have indicated that they believe they do not need to ask individual artists or songwriters for permission to use their music for AI training. This would allow them to include everyone's music automatically and decide on their own how much each artist or songwriter receives. However, some independent companies have said they will let their artists and songwriters choose whether to take part in the AI training deals they sign.
The major labels have also indicated that they do not plan to share income from their AI deals with session musicians. They intend to treat AI income in the same way as money from streaming and record sales.
On the output
Many AI deals that the music industry has signed so far involve products that will also use existing recordings and songs, or an artist’s voice or name, in the content they create. These products may let users remix existing tracks, create copies of artists’ voices or generate tracks ‘in the style of’ specific artists.
Universal Music in particular seems keen to put pressure on AI start-ups and other digital music companies, including Spotify, to focus on AI products that remix existing music or use existing artists’ voices and brands. The resulting tracks would only be shareable within the AI or streaming company's own platform.
If AI output uses an artist’s voice or brand, the assumption is that the artist must give specific permission. This is because record labels do not usually secure rights to an artist’s voice or brand in standard recording contracts. For this reason, the major labels have generally indicated that artists will have to opt in to these uses.
Again, some labels and publishers have started contacting artists and songwriters to discuss whether they want to opt in to AI products that use their music, voice or name in the output. But for the most part, there is still little public information about how these services will work or how artists and songwriters will be paid.
Trends + developments
The lack of transparency, along with suggestions that major labels will not ask individual artists and songwriters for permission for basic AI training, is starting to cause controversy in the music community. Artists, songwriters and their managers are increasingly speaking out, usually through their industry organisations.
Most groups representing music creators are asking for similar things, including:
- Artists and songwriters must always give specific permission, including for basic AI training.
- Labels and publishers must be open about how they will share the money.
- The lower streaming royalty rates often found in recording contracts must not be applied to AI income.
- AI deals must always split income equally between the song and the recording.
- Session musicians must also receive some form of payment.
It is not yet clear how far labels and publishers will agree to these demands voluntarily, or whether major and independent companies will take different approaches. Some people in the music creator community now argue that copyright law must be changed to make sure the whole industry follows the same rules and that all music creators receive a fair share of any money earned through AI deals.
Big debate #03: What to do about AI music
As explained above, the debates about what AI companies, record labels and music publishers must do are far from over. But while these arguments continue, millions of AI-generated tracks are being released and uploaded to music streaming platforms.
Deezer recently said it receives “more than 90,000” tracks made entirely by AI every day. These make up “over 50% of the total number of tracks uploaded to the platform”.
This leads to a third debate: what should we do with music made by AI? Under current laws in most countries, music created entirely by AI does not receive copyright protection. But should it receive some form of protection?
Meanwhile, should streaming services…
- accept tracks made by AI?
- pay royalties for tracks made by AI?
- mark AI-generated tracks clearly on their platforms and in their apps?
- promote AI-generated tracks through playlists and recommendation algorithms?
Should media companies, especially public service broadcasters such as the BBC, play or include AI-generated music in their playlists? Should the industry allow AI-generated tracks to appear in the music charts?
Over the past year, streaming services, media companies and organisations that compile music charts have started publishing and updating their policies on AI-generated music. At the same time, the industry has been discussing and proposing ways to identify AI music and decide how it should be handled when it is distributed and streamed.
For all these questions, an important distinction is being made between AI-generated music and AI-assisted music. AI-generated music is created almost entirely by an AI model, with very little guidance from a person. AI-assisted music is made by a human creator who uses AI tools as part of the creative process.
Most people agree that AI-assisted music should have copyright protection and be treated much like music made entirely by people. By contrast, they believe that music created entirely by AI should face various restrictions. But this raises one final question that no one has really answered yet: where should we draw the line between AI-generated and AI-assisted music?
For many people, including music creators, music fans and others who work in the music industry, that line is not clear at all.
One person might consider a track to be ‘AI-generated’, while another might call it ‘AI-assisted’. Some people believe it is fine to use AI to help create a catchy musical phrase or beat, or to help write lyrics, and that this use does not need to be pointed out. Others argue that any use of AI in a track should be clearly marked so listeners can decide for themselves whether they want to listen to it.
One thing is certain: AI is developing incredibly quickly, and these discussions will help shape the future of the music industry for many years to come.