“Stem-separating” AI is revolutionising the music industry

Associate Jonathan Coote discusses the copyright implications of using AI tools to separate “stems” in music production.

Jonathan’s article was published in World Intellectual Property Review, 26 March 2024, and can be found here.

Background

Many of the headlines around AI and music have been about AI creating new music, or its training on existing music. However, in professional and home studios AI is already being used as a tool by producers and engineers. A lot of these uses are simply improvements on existing techniques, from automated mixing and mastering to sophisticated plugins in digital audio workstations (“DAWs”, like GarageBand or Pro Tools). However, some tools could be game-changing.

Stem Separation

Audacity is one of many companies which claims that its technology can now separate individual “stems” in a recording. “Stems” refer to the different parts of a multitrack recording, typically for each instrument or vocal part. This new technology could mean that you could separate out the funky drums from your favourite James Brown track, or isolate Taylor Swift’s voice. From an acoustic perspective, when tracks are combined, it has been extremely difficult (if not impossible) to separate these interlapping sound waves (think un-mixing the eggs, flour and butter from a cake batter…).

This could completely reinvigorate sampling as a creative tool – with incredible possibilities. But what are the potential copyright implications?

Sample Licensing

Copyright for sampling is notoriously complex, with licences typically needed for both the composition (i.e. the music and the lyrics, aka the publishing rights) and the sound recording rights (aka the master rights), which are usually held by different entities, the publisher and the label, respectively. In addition, if these have not been waived or assigned, you may also need consent for the exploitation of moral rights (for example, the right of paternity and the right of integrity) or performers’ rights and, for vocal samples, it is often good practice to ensure that the original artist has approved the use. Beyoncé found this out to her detriment when her track “Energy” sampled Kelis’ “Milkshake” without Kelis’ consent. Whilst Kelis had assigned away her rights so did not have a legal case, she had a case in the court of public opinion and Beyoncé subsequently removed the sample.

In the UK, to infringe you must copy the whole of a work or a “substantial part” of it. With respect to a musical or literary work, this is a qualitative test and there is no safe number of notes or words you can take. But there are limits. For example, you could not infringe the rights in a composition by copying one hi-hat (or it would have to be the world’s most incredible hi-hat…). Even if you take a tiny sample, this doesn’t mean that you can get away with not paying for a licence, though, as you may still need a licence for the sound recording.

Substantial part of a sound recording – “lend me your ears”

On current interpretation, the definition of a substantial part in relation to sound recordings is much broader, and industry practice is to license samples on the recording side no matter how short. Drawing on EU case law (Pelham v Hutter (C-476/17)), the only exception seems to be when a sample is used in a “modified form unrecognisable to the ear”. In Pelham, the European courts found that the reproduction of a two-second repeated drum loop by Kraftwerk was infringing. The quotation copyright defence was not applicable because the original track was not identifiable and there was therefore no “dialogue” between the two works.

Based on the current law, it is difficult to think of many examples when the use of a small sample of a sound recording would not be infringing. On the one hand, you could try and modify or bury a sample so that it was in a “form unrecognisable to the human ear”, but it is unclear what this means in practice. If someone has identified it, does that mean you per se cannot rely on this exception? Does it only mean an untrained ear? Or could it be the expert ear of a sound engineer? Alternatively, you could “quote” the work such that it was clearly identifiable and there was a dialogue created between the works. Yet, this would then run the obvious risk that you would have to use a distinctive and therefore a recognisable sample, making your potential infringement clear. The chances are that the owner will still demand a licence, no matter how eloquent your dialogue. Lastly, it is possible to imagine a world in which sound recordings could be considered “pastiche”, but, given the difficulties in applying this test so far, this is not an exception which can be reliably argued as of yet.

The current law therefore appears to give sound recordings an elevated status above other copyright works. There are, though, understandable commercial and legal reasons for this:

• the protection of sound recordings is designed to protect investment, and is therefore arguably an economic right rather than a right to creative expression;
• you cannot accidentally sample a recording (as you may be able to do subconsciously if you were to reproduce part of a composition); and
• lastly, a system where you have to get licences for all uses of sound recordings is – for its faults – at least a clear system.

Can you catch them all?

With an influx of stem-separated samples soon to be clandestinely buried in future hits, how will enforcement work?

In practical terms, to sue for copyright infringement you first have to find the infringement. One of the evident issues that such tools raise is that it will become increasingly difficult to hear when a sample is used if it has been taken from an individual stem. However, whilst this may cause potential samplers to rejoice, there is a caveat. As AI stem-splitting tools spread, it is possible that other automated tools will arrive built on similar technology which will conversely allow infringing uses to be detected. This is already the case for images on the internet, and could become more widespread in the music industry.

Liability

In addition to infringing copyright when sampling one track into another, there may be wider questions about the availability of such tools and whether there is liability not only for users, but for the tool providers themselves. Under UK law, there potentially could be an argument that these providers could be secondarily liable or liable as joint tortfeasors for facilitating copyright infringement by their users.

However, in practice, rightsholders are unlikely to take action against many of these tools. In addition to the legal complexities of such arguments, DAWs already include many tools which allow you to conduct all manner of copyright infringement and this has not generally led to court proceedings. Furthermore, it is questionable that the main use of these tools is to facilitate copyright infringement, as opposed to, say, P2P file-sharing sites. Unless rightsholders are seriously concerned about this they will probably not risk the wrath of the industry, with producers and engineers who regularly use these products unlikely to be keen on their creative tools being policed.

If the tools are themselves trained on copyright works, though, this may be a different story…

Conclusion

The use of stem separation software is an incredible creative opportunity provided by AI. However, the ease with which it could allow the circumvention of established norms points to AI’s ability to challenge our fundamental understanding of copyright law.