There have been many discussions over the last few years regarding music and AI – and especially in relation to generative AI models that output music, and which have been trained using existing human-created recordings and songs.These models pose numerous different questions for music creators and the music industry, some legal, some commercial and some ethical.
Some of these questions relate to issues where there is quite a lot of consensus and unity within the music industry, with artists, musicians, songwriters, composers, studio producers, collecting societies, and both major and indie record labels and music publishers, more or less aligned. But some relate to issues where there is a growing divide between different stakeholder groups within the music community.
Given all the many different questions and issues, navigating the music and AI discussions can get confusing. To help, CMU has organised many of these questions and issues into three distinct strands – the big three music and AI debates.
This Friday, CMU’s Chris Cooke will join a panel to discuss music and AI at Un-Convention in Manchester, and ahead of that discussion here he provides an overview of each of the big three debates, running through the key questions being asked, and recent trends and developments.
Big debate #01: The obligations of AI companies
On the input
When training a generative AI model to output music, an AI company needs to copy a large quantity of existing music, both songs and recordings, into a training dataset. It then analyses that data to build its model. Additional copies may be made and manipulated during the subsequent training process.
Copyright law gives copyright owners control over who can copy their works, which means in theory the AI company must get permission from all the relevant copyright owners before making these copies. In the case of a recording of a song, that would include the owners of the separate copyrights in the recording and the song or composition.
However, many AI companies did not secure permission from rightsholders before they trained generative AI models, instead arguing they could make copies of existing content, like music, without rightsholder consent because of ‘copyright exceptions’ or ‘fair use’.
Copyright law in some countries includes an exception covering text and data mining – often called TDM exceptions – which allows third parties to make copies of content for that purpose without getting permission from rightsholders. Many AI companies argue that applies to AI training, so – if they do their training in those countries – they don’t need to get rightsholder permission.
US copyright law has a similar principle called ‘fair use’ – if use of a copyright protected work is ‘fair’ permission is not required. Many AI companies argue AI training is fair use.
Lobbyists for many AI companies have also been pressuring lawmakers in different countries to introduce new copyright exceptions to make it easier to use existing works in AI training without getting permission from the relevant copyright owners or creators.
The music industry is pretty much united in seeking to block attempts by AI companies to exploit existing exceptions or the fair use principle, or to introduce new exceptions. The music industry – along with the wider creative industries – is adamant copyright exceptions and fair use do not – and should not – apply to AI training.
To that end, record companies, music publishers, collecting societies and individual music creators have all filed lawsuits in the US, Europe and elsewhere seeking to demonstrate – by setting legal precedents in different courts in different jurisdictions – that AI companies cannot rely on exceptions or fair use when copying music into training datasets in order to train generative AI models.
Among the AI companies fighting lawsuits filed by the music industry are Suno, Udio, Anthropic and Open AI. All of these cases continue to work their way through the litigation process.
There have also been moves to clarify or change copyright law so that – even if AI companies can exploit exceptions in a small number of countries to train models without getting rightsholder permission – those models cannot then be commercialised in other countries where no such exceptions exist.
On the output
Some generative AI platforms also make use of existing music on the output – for example by remixing existing songs – or exploit an artist’s voice or name on the output – by generating vocals that imitate a performer or allowing users to generate music ‘in the style’ of specific human creators.
Where this is happening by design (sometimes it happens by mistake) most AI companies accept that they need to get permission from rightsholders and/or performers.
They generally accept that when existing songs are replicated or remixed in an AI output, copyright exceptions and fair use do not apply. And if an AI output imitates an artist’s voice or is generated ‘in the style of’ an existing artist, it is likely other creator rights are also being exploited, such as personality or publicity rights, personal data rights and trademarks.
It’s actually not clear precisely how personality, data and trademark rights apply to the use of voice or name on AI outputs, and in the UK personality rights do not currently exist. To that end, there have been proposals in various countries for new digital replica rights that specifically apply to AI-generated outputs, and most of these proposals have been endorsed and supported by the music industry.
However, even where these rights are being clarified and evolved, most AI companies are more cautious when it comes to exploiting works, voices and names within the content that their models’ output.
Trends + developments
The big dispute between the music industry and the tech sector about the legal obligations of AI companies is far from over. However, some tech companies have started to negotiate licences covering AI training and outputs with music industry rightsholders.
In some cases, that’s because the AI company wants to do the right thing. Sometimes it’s because they are also streaming platforms that have existing relationships with the music industry. Sometimes it’s because they want to offer remix tools or voice cloning on the output and therefore recognise licences are required. And sometimes it’s because of legal risks.
Whether AI training is ‘fair use’ under US copyright law – meaning existing music can be used for AI training without getting permission – remains an active and very contentious legal question.
Dozens of lawsuits in the US courts, including those filed by authors, book publishers, newspaper owners and film studios, as well as the music industry, centre on the question of ‘fair use’. In the first two cases to get to trial, against Anthropic and Meta, both involving book authors and the written word, the judges accepted fair use arguments put forward by the two tech companies.
However, in the Anthropic case the judge said AI training was only fair use if the AI company got the original copies of the books it trained on from legitimate sources, and many AI companies have relied on pirated material to build training datasets.
And the judge in the Meta case said an AI company’s fair use defence might fail if a copyright owner can show the outputs of any one generative AI model cause market dilution, impacting on the commercial value of the rightsholder’s works.
So, with music, if tracks generated by an AI model could be shown to significantly reduce the commercial value of the human-created music it was trained on, the fair use argument may not apply.
These mixed-bag rulings have created legal uncertainties for AI companies. Under US law a copyright owner can sue for up to $150,000 for each work that has been wilfully infringed and most AI companies have copied millions of works without licence. This means if any one AI company ultimately loses a copyright infringement lawsuit in the US, it could be forced to pay out a trillion dollars or more in damages.
In that context, it is arguably in the interests of an AI company to sort out copyright licences now to remove that risk. And some rightsholders – including music companies – have been able to use the risk of almost unlimited damages to pressure AI companies into negotiating licensing deals.
Which creates a second big debate.
Big debate #02: The obligations of labels and publishers
On the input
Forcing AI companies to agree to licensing deals to 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 to enter into licensing deals with AI companies to allow training with their catalogues, there are a number of new questions that are immediately posed for stakeholders within the music industry.
- Will labels and publishers seek specific consent from individual artists and songwriters before allowing their recordings and songs to be included in the AI training deals they have negotiated, given AI uses were not anticipated in any old record or publishing deals?
- How will any money the label or publisher receives from an AI company be allocated to individual recordings or songs in their catalogues? With AI training, you could possibly argue that every work is equally important so should get an equal share of the money – but is that the approach the industry is taking?
- Will recordings and songs be treated equally? With streaming, more money flows to the recording than the song, whereas with sync the two distinct rights usually earn more or less the same. Will income from AI training be split equally or in some other way?
- Once money has been allocated to any one recording or song by a label or publisher, what share of that money will be paid to the artist or writer? Record deals often have one royalty rate for streams and record sales and another higher rate for licensing deals – will the higher rate be applied?
- Will session musicians who play on a recording – who usually do not receive royalties from streams or record sales – get a share of this new AI income stream?
Although some labels and publishers have started reaching out to some artists and songwriters – or in some cases their managers or lawyers – there is currently very little transparency on all of this, and no commonly agreed template for how these deals should work. The AI deals themselves are hidden behind NDAs.
The major labels and publishers have indicated that they believe they do not need individual artist or writer consent when it comes to AI training, meaning they can opt-in everyone’s music automatically and then unilaterally decide how each artist and writer will be paid. However, some indies have indicated they will give their artists and writers the option to opt-in or opt-out of any AI training deals they enter into.
The majors have also indicated that they do not intend to share any income from their AI deals with session musicians, treating AI the same as streaming and record sales.
On the output
Many of the AI deals entered into by the music industry so far relate to AI products that will also exploit existing recordings and songs – and/or an artist’s voice or name – on the output, by allowing users to remix existing tracks, generate voice clones and generate tracks ‘in the style of’ specific artists.
Universal Music in particular seems keen to exert pressure on AI start-ups – and other digital music companies, including Spotify – to prioritise AI products that involve remixing existing music and utilising existing artist voices and brands, with the outputs only shareable within each AI or streaming company’s own platform.
Where an AI output exploits an artist’s voice or brand, the assumption is that this will require the specific approval of the artist, because labels do not traditionally secure rights in relation to voice or brand in conventional record deals. To that end, the majors have generally indicated that they will seek artist opt-ins in these scenarios.
Again, while some labels and publishers have started reaching out to some artists and writers to discuss how they might opt-in to AI products that use their music, voice or name in the output, in the main there is currently little transparency on how these services will work and how artists and writers will be paid.
Trends + developments
That lack of transparency – and the suggestion individual artist and writer consent will not be sought for basic AI training by the majors – is starting to cause controversy within the music community, with artists, writers and their managers increasingly speaking out on this issue, usually
Most music creator groups are demanding similar things, including:
- Specific artist and songwriter consent must always be sought, including for basic AI training.
- Labels and publishers should be transparent about how money will be shared.
- Lower streaming royalty rates that are common in record deals should not apply to AI income.
- There must always be an equal split between the song and the recording with AI deals.
- Session musicians should also be compensated in some way.
It remains to be seen to what extent labels and publishers voluntarily meet these demands, and whether majors and indies end up taking different approaches. But some in the music creator community are now proposing that copyright law reform will be necessary to ensure a uniform approach across the industry, and that all music creators fairly share in any money generated by the AI deals.
Big debate #03: How to deal with AI music
As noted, the big debates over the obligations of AI companies – and labels and publishers – are far from over. However, as those debates continue, millions of AI-generated tracks are being released into the world and uploaded onto the streaming platforms.
Deezer recently said that it receives “more than 90,000” fully AI-generated tracks every day, which is “over 50% of the total number of tracks uploaded to the platform”.
This results in a third debate, what do we do with AI-generated music? Legally speaking, does AI-generated music enjoy copyright protection like human-created music? In most countries, as it currently stands, it does not – but should it in some way?
Meanwhile, should streaming services…
- accept AI-generated tracks?
- allocate royalties to AI-generated tracks?
- label tracks that are AI-generated within their platforms and apps?
- push AI-generated tracks through playlists and their algorithms?
And should media companies – especially public service broadcasters like the BBC – play or playlist AI-generated music, and should the industry allow AI tracks into the music charts?
Over the last year, streaming services, media companies and chart compilers have started to publish and evolve their policies towards AI-generated music. Meanwhile there have been discussions and proposals within the industry regarding how AI music can be identified and how it should be treated when it is being distributed and streamed.
When it comes to all these questions, a big distinction is being made between AI-generated music – where the AI model does all the work with minimal prompts from a human – and AI-assisted music – where a human creator uses AI tools as part of their creative process.
The general consensus is that AI-assisted music should be protected by copyright and pretty much treated like entirely human-created music, whereas entirely AI-generated music should be restricted in various ways. But that poses one final question that no one has really answered as of yet: where do we draw the line between AI-generated and AI-assisted music?
For many people, including music creators and music fans, as well as everyone working in the music industry, that line is far from clear.
What might be seen by one person as “AI-generated” could be seen by another as “AI-assisted”. Where one person might say using AI to help create a hook or beat – or to help write lyrics – is OK and shouldn’t need to be highlighted – others would argue that any use of AI within a track should be clearly flagged and that the people listening to the music should be allowed to make up their own minds.
But one thing is certain, the incredibly rapid pace of AI developments means that all these conversations are going to shape the future of the music industry for years to come.
