By: James Miller and Ajay Partap Gill
Canadian copyright law, as embodied in the Copyright Act of Canada, seeks to strike a delicate balance between the rights of creators and the public interest in accessing and using creative works. The advent of AI generated voices in music has challenged the objectives of Copyright law as the possibility of AI musical works which are indistinguishable from those created by our favourite artists has become a looming inevitability. It’s unclear whether the ability of AI-generated musical works to mimic human voices infringes upon the rights of the individuals whose voice is being mimicked which may have significant implications for the promotion of creativity among artists. Questions also arise as to the subsistence of copyright in those musical works which used AI in their creation, can we justify affording copyright protection to those who have used AI to create musical works in view of the overarching objectives of copyright law? There is also significant ongoing discussion regarding copyright infringement in ‘data mining’, the mechanism for gathering data to train generative AI models which allows the close mimicking of musical artists. This post seeks to explore some of these issues and discuss how AI generated musical works which mimic popular artists might interact with the Canadian Copyright regime.
Copyright in AI generated voices
The human voice, in and of itself, is not subject to copyright protection under Canadian Law. This is largely because the law recognizes copyright protection for fixed, tangible expressions of creativity, not the means by which those expressions are created. The case of Gould Estate v Stoddard Publishing Co[1] articulates this principle succinctly, where it was held there is no copyright in oral utterances such as casual conversation. The case reasoned that a claim in copyright in oral statements presupposes that every utterance deserves the valuable protection of copyright, a notion which runs contrary to common law’s development of copyright. This principle really goes to the core of why the realism of AI generated music is a particularly scary prospect for artists.
It’s for this reason we’re able to enjoy the work of celebrity impressionists who aren’t in violation of any copyrights even if their impersonations are truly spot on. The use of AI generated voices in itself, is not a violation of copyright, which is why right now the key question as it pertains to copyright infringement and generative AI is the data used to train them and how it may qualify as fair use/dealing.
Subsistence of Copyright in AI Works – who’s the author?
The question of who the author of AI generated works is a question that admittedly, feels straight out of a Sci-fi movie. As it stands right now, Generative AI can’t be given property rights and at least at it’s current level of sophistication is more like a highly advanced tool capable of spitting out detailed outputs based on it’s wealth of training data, rather than an autonomous intelligent being that we associate human authorship with. Sorry Sci-fi enthusiasts.
At it’s core, for generative AI to properly fulfil it’s purpose to any meaningful degree, whether it be for text, images or music, a meaningful degree of human intention is needed to guide the AI. It is a requirement in Canadian Law that works have a sufficient degree of originality in order for copyright to subsist in it. Under leading case CCH v LSUC[2], an exercise of skill and judgment that is more than a purely mechanical exercise is needed for a work to meet the originality requirement. As it pertains to AI generated musical works, this requirement will likely depend on a case by case basis and will depend much on how the user actually interacts with the AI.
Widely popular generative text AI Chat GPT provides an illustrative example. It is one thing if I ask ChatGPT to write me a poem in the style of Shakespeare. This exercise is relatively mechanical and really doesn’t require much effort on my part. ChatGPT in response would likely generate a poem in Shakespearean English with some general stylistic similarities to the works of the great English poet. It is a different thing if I give ChatGPT extensive instructions on the subject matter of a poem I want created and the themes I want to include in this poem, as well as specific instructions as to style such as rhythm, metre and rhyme. This exercise clearly requires a degree of skill in the way I instructed the AI and a demonstration of judgement through my exercise of choice in, what features I specifically wanted in my poem.
The US copyright office seems to follow this view, issuing advice last month stating that copyright protection can only be afforded to AI generated works so long as it reflects the author’s ‘own mental conception’[3]. The advice issued builds on its rejection of copyrights in images featured in the comic book ‘Zarya of the Dawn’ made by artist Kris Kashtanova with the help of generative AI MidJourney[4]. The approach adopted by the US copyright office seems to fit neatly within the Canadian conception of authorship and is likely to be persuasive authority,
The same reasoning can be applied to the creation of musical works, the key question as it pertains to authorship, is what extent the user instructed the AI to create the end product. In what has become a recently viral case in AI generated music, the user by the name of ‘Ghostwriter’ posted a song titled ‘heart on my sleeve’ which he disclosed as an AI version of a Drake x The Weeknd collaboration. The song quickly went viral across different social media platforms but was shortly taken down by Universal Music based on the use of an unauthorised sample since the start of the song featured a sample of producer Metro Boomin’s famous producer tag[6]. Notwithstanding the sample use, it’s unclear to what degree Ghostwriter used generative AI to create the song. If he was truly the creator of the entire musical composition and merely used AI generated voices to mimic Drake and the Weekend to perform lyrics which he wrote, it safe to say he is the author of ‘heart on my sleeve’. In such a case the use of AI generated voices is no different to using an instrument and there is clearly an exercise of skill and judgement in creating the musical work as a whole. There exists musical generative AI’s capable of creating royalty free music from a textual description of the music you want made (i.e. Mubert)[7], which is akin to ChatGPT in the way the user interacts with the AI interface. It’s unclear how specific your instructions to such an AI would need to be in order to warrant copyright protection and these are questions which are ultimately untested by Canadian Courts.
Copyright infringement in Data Mining for Training Data
One of the legal challenges with AI generated music is that the AI systems are trained on content (I.e. songs and lyrics) that are protected by copyright. The training of Generative AI systems to produce music might infringe the copyright owner’s rights unless it falls under doctrines of fair dealing (in Canada) or fair use (in the USA)[8]. These doctrines create an exemption for certain unauthorized uses of copyrighted materials. We will focus our discussion on fair dealing as we are focusing on Canadian law.
The fair-dealing exception (s.29) of the Canadian Copyright Act allows use of copyright-protected materials under certain contexts. There is a 2 step test outlined in CCH for fair dealing. First, does the purpose of the dealing fit into one of the listed purposes? Second, was the dealing fair? Let’s look at how data mining fits with these 2 steps of the fair dealing test.
Step 1: Does the purpose of the dealing fit into one of the listed purposes?
Under the first part of the CCH test, the purposes under which fair dealing can be found are research, private study, criticism, review, education, satire, parody, and news reporting as long as what is done with the work is “fair”[9]. We will explore the purposes of research and parody as it seems these are the 2 categories under which data mining copyrighted music in the context of training generative AI models would likely fall under.
Research
Research is one purpose under which data mining to train AI generative music models can potentially be seen as fair dealing. CCH holds that “The fair dealing exception under s. 29 is open to those who can show that their dealings with a copyrighted work were for the purpose of research or private study[10]. ‘Research’ must be given a large and liberal interpretation in order to ensure that users’ rights are not unduly constrained[11]. Additionally, research is not limited to non-commercial or private contexts, and probably does include commercial research[12]. Similar to how OpenAI classifies ChatGPT as a research tool, if AI music generators classify their tools as research this may open the door to them as falling under the definition of research. However, if the music is generated primarily for commercial purposes then it may be difficult to argue that it is for research purposes. Whether AI generated music classifies under research will be a complex question that courts will have to settle in the near future.
Parody
Parody is another category under which AI generated music could fall under. Per United Airlines v Cooperstock, the two basic elements to establishing parody are 1) the evocation of an existing work while exhibiting noticeable differences and 2) the expression of mockery or humour[13]. This is another likely category AI-generated music can potentially fall under as it will be similar to an existing work but have different lyrics. However, the question of whether AI generated music is an expression of mockery or humour will be an open question for courts to answer.
If it is established, AI generated music falls under one of these categories then we must look at whether the dealing was fair.
Step 2: Was the dealing fair?
To determine whether a dealing is fair or not, we must look at the various factors laid out in CCH: the purpose of the dealing, the character of the dealing, the amount of the dealing, the nature of the work, available alternatives to the dealing and the effect of the dealing on the work[14]. Below, we will analyze how each of these factors may apply to AI generated music and the further questions this might raise.
- The purpose of the dealing looks at the user’s motive[15]. The purpose for which a user is making AI generated music will be a key factor in whether it counts as fair dealing. If it falls under parody or research as described above, then this factor may be satisfied. Given the potential for users to potentially commercialise their AI generated works by selling it on streaming or selling their AI created songs, Canadian courts may question whether use of copyrighted works as training data fits the purposes of research or parody at all. Only time will tell how Canadian courts will answer this question
- The character of the dealing looks at what was done with the copyrighted work[16]. This includes whether it was used on an isolated or ongoing basis and how widely the work was distributed. Often a copyrighted work, like a song, is only used once to train a Generative AI model which may point to fair dealing[17]. Additionally, destroying the training data, which includes these songs, after they have served their intended purpose of training the model may point to fair dealing. However, on the flip side, once trained the actual model is further fine tuned and widely distributed to generate various outputs/songs. Therefore this is a complicated factor which does not clearly point to whether the dealing is fair or not and will have to be determined by courts on a fact-specific basis.
- The amount of dealing looks at how much of the work was used and how important was the content that was used[18]. Typically numerous full songs are used to ensure the model has a wide variety of input. However, the exact number of songs used in training is not currently clear as these AI-generated songs do not publicize their training data. Record labels and artists will likely argue this is not fair dealing if a large part of an artist’s music catalog is used as training data since this is a substantial part of their copyrighted work and they want to be compensated for it. Therefore, this is another factor that will have to be determined on a case by case basis, depending on how much of an artist’s catalog is used as training data.
- Alternatives to the dealing discusses whether alternate methods could be used to achieve the dealings ultimate purposes whether it be through the use of alternative methods to train the AI which are deemed more ‘fair’ or using alternate non-copyrighted works entirely[19]. To create a robust and accurate model, generative AI requires a vast array of data. In the context of musical generative aI, this might entail using an artist’s entire discography to create an AI voice which actually mimics a given artist’s voice. Record labels may take particular issue to this but it’s unclear whether there are any alternatives. An argument could be forwarded that in order to achieve its ultimate purpose of mimicking an artists voice, whether it be through the guise of parody or research that, using an artists entire catalogue as training data is necessary, and alternatives would limit the realism of the generative AI’s end product.
- The nature of the work being used looks at whether dissemination of copyrighted work aids the public interest[20]. On one hand, these models may empower small independent artists, who often lack resources, to use an AI-generated model to include a feature on their song for an artist they could never afford to pay for. However, on the other hand, record labels will argue this violates their economic interests and they should be entitled royalties for the use of their works by smaller artists. This will be a highly debated factor that will involve balancing the economic interests of copyright holders, such as artists and record labels, with empowering artistic expression for smaller artists.
- The effect of the dealing on the original work examines whether the use competes with the market of the original work[21]. This is a highly complicated question as it remains to be seen how AI generated songs affect the music industry. Currently these AI generated works are labelled as being “AI generated” (i.e. AI Drake) and do not pass on as the artist themself (i.e. Drake). It will be interesting to see whether a new category develops for AI generated music itself or if the AI generated music becomes so good that they are competing for streams with the artists themselves.
The AI generated song “heart on my sleeve” featuring AI versions of Drake and The Weeknd was streamed hundreds of thousands of times on Spotify before it was taken down showing the viral potential of AI-generated music to compete for streams[22]. If AI generated music starts to compete with original artists themselves and is shown to actually hurt their streams and sales, this may be a significant factor towards fair dealing not being found. Further, the question of whether artists are owed royalties for AI-generated music will be a hotly debated topic within the music industry. On one hand, AI-music generators may argue that their AI-generated music is creating more buzz and exposure for an artist like Drake. However, record labels will argue that this AI-generated music is using their name and goodwill to compete with their copyrighted works for streams.
This raises even more questions for dead artists and the implications of using their voice in AI-generated music. Imagine if someone is able to feature Frank Sinatra on a modern day hip-hop song using an Generative AI model. This may have the effect of drawing more publicity for deceased artists and could revamp the music of generations passed to a younger audience. Only time will answer these questions but the actual effect that AI generated music has on the sales of the artists they are mimicking will have a significant impact on the likelihood of a finding of fair dealing.
This analysis of whether the data mining of copyrighted training data to train Generative AI music models constitutes fair dealing or is copyright infringement shows that there is currently no clear cut answer to these legal questions. Courts will have to balance the economic interests of copyright holders with the public interest of artistic and creative expression.
Although the specific issue of whether using copyrighted songs and lyrics to train Generative AI models that produce music counts as fair dealing is one that has not been answered by courts yet, copyright infringement cases related to the use of generative AI are evolving rapidly. In January 2023, Getty Images sued the Generative AI art company, Stable Diffusion, in the UK, alleging that its tool Stability AI unlawfully copied and processed millions of images protected by copyright to train its software program[23]. What the court rules in this case will likely be a persuasive authority in Canada that will pave the way for cases involving the use of copyrighted works to train Generative AI music models.
Footnotes:
1. Gould Estate v Stoddart Publishing Co Ltd (1998), 39 OR 555
2. CCH Canadian Ltd v Law Society of Upper Canada [2004] 1 SCR 339 [CCH]
3. U.S. Copyright Office ‘Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence’ (16 May 2023) online: <https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence>
4. Blake Britain ‘US Copyright Office says some AI-assisted works may be copyrighted’ (15 March 2023) online: Reuters <https://www.reuters.com/world/us/us-copyright-office-says-some-ai-assisted-works-may-be-copyrighted-2023-03-15>
5. Nilay Patel ‘AI Drake just set an impossible legal trap for Google’ (19 April 2023) online: The Verge <https://www.theverge.com/2023/4/19/23689879/ai-drake-song-google-youtube-fair-use>
6. Ibid.
7. https://mubert.com/render
8. Chris Hunter et al., “Does Generative AI Need To Infringe Copyright To Create?” (16 March 2023), online: Mondaq <https://www.mondaq.com/canada/copyright/1294336/does-generative-ai-need-to-infringe-copyright-to-create>
9. Dan Brown, Lauren Byl, Maura R. Grossman (2021). Are machine learning corpora “fair dealing” under Canadian law?. UWSpace. http://hdl.handle.net/10012/17708 [Brown]
10. Ibid.
11. Ibid.
12. Ibid.
13. United Airlines v Cooperstock 2017 FC 616
14. CCH, supra note 2
15. Ibid.
16. Ibid.
17. Society of Composers, Authors and Music Publishers of Canada v. Canadian Assn. of Internet Providers, [2004] 2 S.C.R. 427, 2004 SCC 45 [Socan]
18. CH, supra note 2
19. Ibid
20. Ibid
21. Ibid
22. Samantha Murphy Kelly, “The viral new ‘Drake’ and ‘Weeknd’ song is not what it seems” (19 April 2023), online: CNN <https://www.cnn.com/2023/04/19/tech/heart-on-sleeve-ai-drake-weeknd/index.html>
23. James Vincent, “Getty Images is suing the creators of AI art tool Stable Diffusion for scraping its content” (17 January 2023), online: The Verge <https://www.theverge.com/2023/1/17/23558516/ai-art-copyright-stable-diffusion-getty-images-lawsuit>
Communications Law