Hi, everyone. My name's Alex, and this talk's going to give a snapshot of where AI is being used in the music industry today, and hopefully where it's going next.

So what is the music industry? So it's a little hard to define concretely, but broadly speaking, it's individuals and organizations that either create, distribute, and sell various forms of music, put on performances, so venues or performers, or aid, train, assist, represent, and supply music creators.

The music industry is not a static thing and it's changed a lot over time. It generally changes with technological development.

So we're in the fourth industrial revolution alone, often called Industry 4.0. So these are technologies like AI or cloud computing or robotics, all these kinds of things.

And with this kind of new technology, there's a lot of innovation going on in music tech. I'm sure a lot of you have seen products like maybe Suno or Yudio, these AI music generators.

At the start, we had some AI-generated music, I think. And generally, there's a lot of tension between these kind of music technology companies and the pre-existing music industry, particularly like music publishers, like big record companies like Universal Music Group or Sony, etc.

So where is AI currently being used in music? So a couple of different areas.

I imagine if you say AI in music to most people, they'd probably think of generative music, but there's a lot of other use cases beyond this. So for instance, there's singing voice cloning.

Maybe some of you heard the Drake AI, that song came out earlier in the year. source separation, so splitting music into its separate instruments, mastering perfect music, writing lyrics, creating artwork, public relations and marketing, and even venue security and management.

So if you go to the O2 Arena in London, they actually use AI face detection now when you're walking to the venue for security purposes. So it's being used in a lot of places that maybe you wouldn't immediately think of.

Why is all this research in AI music going on? There's a lot of potential benefits that come out of it.

From the creative side, there's further accessibility to the arts. Anyone in this room could make a song using Suno or Yudio.

It's the chat for music, essentially. There's been all kinds of new job opportunities from this.

Developing AI music models, for example. Increased productivity, we've seen some of that today.

New methods of fan engagement, so fans can now make songs in the style of their favorite artists. And artists can use AI to also speak to fans. New revenue streams, and of course the new creative tools and mediums of expression.

So all great stuff.

But there's also a lot of risks and issues that come with using AI in the music industry. So automation and job displacement is a big one.

A lot of artists are worried about their jobs, but also, as we've seen from the other examples, it's not just artists, it's anyone who works in the music industry, really. There's a lot of discussion about how to use training data ethically and who kind of owns AI generated music and if we create AI generated music how do we fairly compensate either the people who created that model or the artists who provided training data for it. If this isn't handled well, this has led to cases of data poisoning where artists have sabotaged their own work to make it unpalatable for AI training.

We've got de-skilling. If you're replacing artists, you're actually losing skills. This isn't unique to music, this is in a lot of industries.

Great for stifling AI, reproducing AI models, and of course the huge environmental impact, mainly from energy costs of training and using these AI models. And also all the kind of rare materials to make the specialized computer chips.

It's a huge impact. And the big picture.

There's been a couple of recent reports that I think really paint how difficult this is maybe going to be for the industry to change in the next few years. A recent report said music industry workers are estimated to lose a quarter of their income to AI in the next four years without intervention.

These are workers that are already pretty precarious in a lot of cases. The UK government estimates that between 10% to 30% of jobs in the whole UK economy could potentially be automated with AI.

And this disproportionately affects the creative industries in the UK. And also, you can't really restrict AI anymore. All you need to run it is a computer and internet. It's here.

So what do people think about AI in the music industry?

So first we're going to see what kind of musicians and creators are thinking about it. So there's these two surveys which I'm taking this data from and about one in four creators are currently using AI in some form. Some of them are using it to create artwork for their releases, or to create assets or engage with fans, and of course in music production, including 3% of people use it to create entire songs.

But we can also see that the majority of creators are not using AI at all, and one of the main reasons for this is artistic and creative reasons, quality of models, although this is getting better all the time, and copyright concerns.

So what does the music industry think? Well, there's been quite a lot of campaigns and lobby groups that have been trying to protect the existing music industry. So in the United States, we have things like the Human Artistry Campaign. We have AI for Music, which is like a conglomerate of music technology companies saying human centricity should be key. to any AI in the music industry.

We also have some interesting court cases at the moment, so the big music labels are suing these AI song generators, Suno and Udio, in the United States, because these companies basically admitted to scraping any music data they can lay their hands on online to train these models.

And there's actually something similar going on in the UK at the moment where the stability AI who makes stable diffusion, they're being sued for a very similar reason. Yeah, and almost like, well, at the time I made this, 40,000 professionals in the creative industries have also backed the statement saying, you know, don't train on our data without our consent. So it's quite a divisive issue.

And even for people who aren't directly involved in the music industry, just general music consumers, there's also some quite strong opinions that generally point that AI should be regulated to some extent, particularly in the UK.

So this is based on two surveys. One is kind of UK specific, the other is more global and we can see like majorities things, you know, going should play a role in setting restriction of what AI can do.

Lots of concerns about, you know, is what's AI being trained on, are the asks being fairly compensated? And people are even worried about listening to AI generated songs without realising.

This is starting to make some headway. I mean, if any of you use Instagram, you see there's the something like generated with AI label that you can add to posts. But this kind of thing doesn't exist for music yet, and maybe it's on the way.

So there is some regulations starting to emerge. So the kind of G7 nations have signed something called the Hiroshima AI process, and this is to generally restrict, to develop AI in a safe way, but it doesn't really have any legal backing, it's just high-level principles. But if you go to specific countries, a regulation is starting to emerge. And these generally, in the kind of Western countries, they're trying to balance between the fairly strong creative industries, like in the UK, the creative industries are the third biggest contributor to GDP, for example, and enabling technology companies to innovate.

So if we take a couple of case studies, the European Union, the United States, and the UK, the EU currently has the most extensive regulation on AI. They've introduced their AI Act, which will be coming into power, into force in the next year or two, with some parts of it already applicable. So this forces companies developing AI to, for example, label AI-generated content and publish summaries of copyrighted training data.

It's good, but there's also a lot of limitations. People say it doesn't go far enough. There's, for example, a lot of self-assessment involved and a voluntary approach to environmental considerations.

The US is starting to get there, although it's obviously in a bit tumultuous political situation currently, but it's got a patchwork of national and state level regulations, so existing copyright and personality rights. Quite interesting, the first legislation was Tennessee's Elvis Act, so I don't know how many people know this, but Tennessee is where Elvis is from, so to protect their tourism industry, they've introduced this bill, that makes it illegal to make AI voice clones of celebrities without their permission.

And there's also a few other bills that are kind of under consideration called the No AI Fraud and No Fakes Act, which, again, more legislation on the creation of deep fakes. They always come up with very creative names for their bills, which I have to give credit to.

And the UK is probably the furthest behind at the moment. It doesn't have any explicit AI legislation, but it does have the Copyright Act of 1988, which actually, considering how old it is, it does explicitly allow for computer-generated works, which is quite interesting.

But it has a limitation that is kind of ambiguous in that If you make computer generated work, you kind of have to fight out in court for who actually made it. Was it like the artist? Was it the person who made the AI model? The data controller? The organization? An investor? It's all kind of up in the air.

So there's a lot of lobbying to maybe change this at the moment.

And there's also been kind of a political football of this text and data mining exemption, which maybe some of you have heard of, where AI companies have been lobbying so they can basically train on any copyrighted data. Whereas obviously like creative industries, they want this, they want to be fairly compensated for it. So yeah, it's an interesting and developing situation.

To round off the talk, I'm going to make a couple of recommendations of where I think the music industry should go. In 2021, the European Commission proposed this concept called Industry 5.0. So we've talked about Industry 4.0 technologies already, and the general aim of these is kind of increasing production efficiency, flexibility and upskilling. But the difference with Industry 5.0 is that these technologies shouldn't just do these things above. They should also achieve societal goals.

And so the Industry 5.0 concept or framework or however you want to call it, kind of has three main principles and these are sustainability, resilience and human centricity. So the recommendations I make are kind of in line with this framework.

So just a quick comment on, so a kind of practice for the use of copyrighted data that's currently emerging in the music industry, or that they seem to like, is what's called a dynamic attribution model. So attribution, this means you're paid according to the percentage of data that is yours that's used in a training set to train an AI model, or the percentage of your data that's used in the output of an AI model, but that's kind of much harder to measure at the moment. You have opt-in so basically the people who's training data using they have to explicitly say you can use my data for training It can't just be taken like the text data mining exemption we're talking about and dynamic so this means you can give or remove your permission at any time and

So to make a couple of recommendations, I'm sure there's quite a lot of different professions in here, so there's a few recommendations for everybody.

So if you're an AI practitioner, I would recommend adopting ethical development guidelines and to discuss and disclose ethical issues. There's a couple of good guidelines. that I would recommend.

So for instance, AI for music we talked about earlier if you're a music technology company. But there's also the machine learning technology readiness level framework, which assesses the market readiness of machine learning models. And this also takes kind of ethical issues into account. So this is a really good framework to look at.

At the model stage, you should design efficient models and disclose environmental impact. So there's all kinds of ways to reduce the model framework and to measure the carbon emissions of your model. So I would recommend looking at these.

You should support the industry grassroots. Everything from creators to early career professionals to ensure the human centricity of the industry and maintain a creative talent pipeline.

You can't train AI without human data. So you need to support this. And yeah, general advocate for fair legal frameworks.

If anybody's maybe looking for the next big thing, there's some good commercial opportunities that kind of could support industry 5.0. I think obviously supporting the dynamic attribution model.

We talked about a couple of companies in this space already, Landa, Surreal, Prorata. But there's definitely space for more companies to grow in this space.

And also licensing and environmental compliance. This gets really tricky with the different legislation in different countries.

So particularly doing this internationally, I think helping artists and companies to use data and use AI responsibly is, I think, could be a good opportunity.

And there's also, there's still loads of active research going on in AI music. So actually, in Queen Mary, there's about 70 PhD students researching AI music, and it's all in different things.

But there's a couple of open questions that I think are worth looking at, including detecting deep fakes, whether this is singers or AI-generated music. There's actually someone in the room today who's looking at this. Sad over there.

There's assessing musical similarity when you're generating music and there's also creating and identifying watermarks in AI generated music to help with compliance and IP enforcement.

Great, so just coming to the end, I just want to say thanks to the people who funded this research, so UKRI and Queen Mary. Also thanks to Lord Tim Clement-Jones for having a chat with me about AI and the creative industries in the UK. And yeah, thank you for listening.

So this talk was based on a paper which I'm presenting at the AAAI conference next week. You can check out the paper, it's already there. I had to kind of skim over a lot of the details with this talk, so I would really recommend giving it a read.

Further Discussion

But also, you can ask me some questions now, or have a chat to me in a bit, and that would be great. Yeah, get in touch here. Thank you.