Generative Music's Not So Fair Use
Should the unlicensed use of music copyright be protected under fair use?
Fair use - the use of copyrighted material without the permission of the rights holder - is determined using four pieces of criteria:
The Purpose and Character of the Use: Is the use for nonprofit or commercial purposes? Naturally, not-for-profit uses have more leeway seeing as they aren’t capitalizing from the works. Does the use add new expression or meaning to the original work (ie. is it transformative)?
The Nature of the Copyrighted Work: Is the original work factual or creative?
The Amount and Substantiality of the Portion Taken: How much of the original work is used, merely a snippet or the whole work? How important is that portion to the work as a whole?
The Effect of the Use upon the Potential Market: Does the use negatively impact the market for the original work?
These factors are considered collectively and on a case-by-case basis, meaning no single factor is determinative.
Let’s break each down in the context of the burgeoning generative music market.
(1) Purpose & Character: All three leading generative music platforms - Suno, Udio, and Stable Audio - have clearly defined subscription models in place, and this stands to reason given the substantial venture backing of the companies. Only Stable Audio has been transparent about its training methods, using exclusive data from AudioSparx, and giving artists the ability to opt out. Suno is a different story; one of its investors admitted in a March Rolling Stone article that he wouldn’t have funded the company if they had licenses. Perhaps the situation has changed since, but the company hasn’t publicized it if so. Similarly, Udio is a black box and has made no public mention of licenses, which is especially disappointing considering the financial backing from prominent artists.
A commercial model with no compensation back to rights holders of the original works? Strike one.
As it relates to the “transformative use” component of the first criterion, it is the intention of generative music companies to create entirely new works. Intention and execution are different, however. In the case of both Suno and Udio, numerous outputs have demonstrated an uncanny resemblance to commercial tracks that they might not have licenses for. These instances are thoroughly detailed in exposés by Ed Newton-Rex (Suno, Udio).
But what about the outputs that are entirely new generations? Transformative use is perhaps the only criterion that could be used to support a fair use argument for unlicensed music. But it alone is not enough.
Whether the data is licensed or not, all generative music companies must have robust content ID systems in place to ensure output isn’t too duplicative of its training material. If the data is unlicensed and outputs are repeatedly duplicative of that data, then that company should fail the transformative use test.
(2) Nature: Music is inherently creative which means fair use is less applicable. Non-fiction information is more protected under fair use “because the dissemination of facts or information benefits the public.” Operators of large language models (LLMs) might thus have a case for not licensing at least some of the nonfiction data used for training.
Recent licensing deals suggest that data sources from the middle of the creativity spectrum - ie. journalistic content on Wall Street Journal - would not be protected under fair use. Just last week, OpenAI struck a 5-year $250mm licensing agreement with News Corp for access to all of its publications. OpenAI struck similar agreements with the Financial Times, Axel Springer, and Reddit. The company also remains embroiled in a lawsuit with the New York Times over alleged unlicensed use of NYT content.
If the use of journalistic reporting data isn’t protected under fair use, then unlicensed music, which is irrefutably more creative, shouldn’t be either.
(3) Amount & Substantiality: AI models are trained using full tracks, not snippets, samples, or stems. This is essential so the model comprehends and is able to recreate a full composition. In some cases, stems are useful to reinforce learning of specific instruments. But training a model using just stems or snippets would yield generations that are disjointed and incoherent.
As such, it stands to reason that the unlicensed use of music for training purposes would fail this test.
(4) Effect on the Market: I believe generative music will amplify human creativity in ways we can’t even comprehend yet. It will complement existing artists and empower new ones. It can be used to sharpen and expedite the creation process, quickly transforming a concept into a track that can be refined in a DAW.
But it’s naive to think that this new technology doesn’t pose competitive threats to the industry. We have already seen numerous instances of AI-generated tracks diverting royalties from deserving artists on DSPs. You can imagine a scenario where a songwriter opts to use an AI-produced track rather than paying a producer. Further, the creation velocity and dramatic reduction in production costs have implications for the sync and production music markets.
Given the implications for the industry at large, it stands to reason that the unlicensed use of music for training purposes would fail this test as well.
Summary
Proponents of fair use in the context of generative music often cite the similarities between human and AI learning as defense of their rationale. After all, the AI model isn’t storing the tracks for re-use, it’s merely learning structure. The argument continues, it would be infeasible to assume that human artists should license all of the tracks that inform and inspire their creations. So why should AI models be required to?
First and foremost, comparing a human to a machine that is fine tuned and force fed volumes of data is incongruous, irrespective of the similarities between a neural network and human brain. We’re talking about an ‘artificial’ machine that is trained for a sole purpose…
Second, generative music is a focused medium which makes a licensing framework entirely feasible. Contrast it to LLMs which require an extraordinary breadth and perpetual feed of information to be the most useful in an increasingly competitive market. This presents an unwieldy licensing conundrum that resembles a never-ending hamster wheel.
Generative music models can be trained with a more targeted data set. They don’t require every new song release to be useful. Of course as music tastes evolve and new genres emerge, it is helpful to refresh and rebalance the training data. But generative music models don’t face the same pressure as LLMs to stay up to date with every piece of new data. As such, there is no excuse for using unlicensed data.
Silicon Valley’s “move fast and break things” and “ask for forgiveness, not permission” mantras are especially inconsiderate when it’s the collective livelihood of artists they are neglecting by using unlicensed music to train their models.
Sources
Measuring Fair Use: The Four Factors
Yes, Udios' Output Resembles Copyrighted Music, Too - Music Business Worldwide
AI in the Music Industry: Fake Streams and Streaming Farms - Music Business Research
OpenAI, WSJ Owner News Corp Strike Content Deal Valued at Over $250 Million - Wall Street Journal