AI Music Is Not Art. It's... The Opposite

A look at legal and ethical issues with AI-generated music.

Image source

In this week’s post, we will take a look at two companies, Suno and Udio, that are caught in an arm-wrestling match to be the best tool in a strange niche of AI; generative AI for music creation.  

As an engineering feat, the mere possibility of AI-generated music is incredible. But will people be willing to pay for it? Is it even desirable for people to create their own music with a few clicks? Suno’s co-founder Mikey Shulman, who is interviewed by the Rolling Stones certainly thinks so:

“(..) Mikey Shulman, a boyishly charming, backpack-toting 37-year-old with a Harvard Ph.D. in physics, envisions a billion people worldwide paying 10 bucks a month to create songs with Suno. The fact that music listeners so vastly outnumber music-makers at the moment is “so lopsided,” he argues, seeing Suno as poised to fix that perceived imbalance.”

Personally, I don’t see Suno, nor Udo, making a dent in the music-listening habits of consumers. Beyond concerns pertaining to copyright, I don’t think the ability to create AI-generated music is valuable or fun enough to pay $10 a month for, not for professional musicians, and even less for music listeners. As I will argue in this post, the notion that AI-generated music or synthetic content can have artistic value is backward.  

The Early GenAI Models for Music 

About four years ago, OpenAI released its text-to-music model, Jukebox, accompanied with code and weights. The release was part of the company’s early efforts to push the boundaries of generative AI at a time when OpenAI was a research-first organization that prioritized science over profits. Three years later, Meta released Audiocraft and Stability AI released Stable Audio, like Jukebox, available for free with code and open weights.  

Google released MusicLM in January 2023. It was arguably the first AI model that could generate high-fidelity songs with complex compositions based on text prompts. To begin with, Google decided to not release the model, in recognition that MusicLM would “reflect the biases present in the training data” and was “raising concerns about cultural appropriation”. Only four months later, Google pivoted on this decision by publicly releasing MusicLM in a test environment called the “AI test kitchen“. The latest upgrade is called MusicFX. 

In November 2023, Google's sister company DeepMind rushed in where Google had originally feared to tread by announcing the release of an advanced music generation model, Lyria, in a partnership with YouTube.

Lyria is designed to leverage AI in music-making with two experiments:

• Dream Track: Lets users create 30-second soundtracks for YouTube shorts with AI-generated voices and musical style of a bunch of popular (and consenting) artists. So far, it's been released in limited beta. 

• Music AI tools: a suite of tools for music creators that can for example be used for “singing a melody to create a horn line, transforming chords from a MIDI keyboard into a realistic vocal choir, or adding an instrumental accompaniment to a vocal track”. The tools will be released later this year. 

Music generated by MusicFX and Lydia is watermarked with DeepMind’s tool, SynthID, so its origin can be detected if need be. The watermark is embedded in audio waveform that is inaudible to the human ear. 

Suno and Udio

Over the course of late March to mid-April, the buzz surrounding generative AI tools for music reached new heights as two companies, Suno and Udio, came out of stealth with impressive products that raised the quality bar for synthetic music a notch or two. 

First came Suno, a company founded in Cambridge, Massachusetts, by four machine learning experts, Mikey Shulman, Georg Kucsko, Martin Camacho, and Keenan Freyberg. The founders were former colleagues at Kensho, an AI startup for financial data that exited with an acquisition by S&P Global (Axios). 

On March 21,  Suno released v3 and announced that its new model could produce “radio-quality music” and songs of up to two minutes in length in a few seconds. Suno is also integrated as a plug-in to Microsoft Copilot. 

Udio released its product on April 10, not long after its seed funding round in December where it raised $10 million from investors such as a16z, Instagram co-founder Mike Kreiger, and the musicians will.i.am and Common. Four out of Udio’s five founders, David Ding, Conor Durkan, Charlie Nash, and Yaroslav Ganin are former Google DeepMind employees, except for Andrew Sanchez who is a former product manager and Doctor of Philosophy from Oxford University. 

Comically, Udio has been referred to as the “Suno killer”. Rolling Stones said:

“the two companies’ output seems closely comparable, though some early users have suggested that on average, Udio‘s output may sound crisper than Suno‘s, with less of the sonic fuzziness that can betray tracks’ machine-created origins."

Others, such as Morning Brew dissent on this opinion but overall, Udio and Suno are strikingly similar.

Udio believes that “AI has the potential to expand musical horizons and enable anyone to create extraordinary music”, whereas Suno’s mission is “building a future where anyone can make music”. 

Both services are currently in public beta and are free to use. Suno offers a Pro Plan that costs $8/month for users who want to generate more than 10 songs a day, and a Premier Plan that costs $24/month for users who want to generate between 500-2000 songs a day. Udio offers a Standard plan for 10$/month and a Pro plan for 30$/month with respectively 1200 and 4800 credits, each credit corresponding to a 1~30 second audio track.

Another commonality between Suno and Udio is that they have so far refused to disclose their sources of training data. On this background, we can ascertain – without a shadow of a doubt – that they have been trained on very large amounts of copyrighted music. That is a problem and against the spirit of an open letter published by Artist Rights Alliance on April 1 and signed by more than 200 musicians, including Billy Eilish, Stevie Wonder, Norah Jones, Jon Bon Jovi, R.E.M, and the estate of Frank Sinatra:

“Some of the biggest and most powerful companies are, without permission, using our work to train AI models. These efforts are directly replacing the work of human artists with massive quantities of AI-created "sounds" and "images" that substantially dilute the royalty pools that are paid out to artists (..)              

We call on all AI developers, technology companies, platforms, and digital music services to pledge that they will not develop or deploy AI-music generation technology, content, or tools that undermine or replace the human artistry of songwriters and artists or deny us fair compensation for our work." 

Honestly, I am a music-lover at heart, and I deeply sympathize with Artist Rights Alliance’s cause. 

Suno and Udio are clearly competing with musicians for the very little profit that is left after the streaming platforms and record labels have taken the lion’s share of revenue. Besides, almost all of the profits on streaming platforms go to the 0.01% of top-trending artists, and AI-generated music will unevenly impact the remaining 99.9%.

Neither Suno nor Udio allow users to generate songs based on the likeness of artists (e.g. "make a song in the style of Eminem") but workarounds are more or less possible. Contrary to Google’s MusicFX and Google DeepMind’s Lyria, Suno and Udio do not have a watermarking system in place so the AI-generated content can be detected.    

Interestingly, Pro and Premier users of Suno are granted the commercial rights to songs which means that the users can monetize their songs on YouTube, Spotify, and Apple Music, and have the right to license the songs out to third parties. Udio is more vague about ownership rights for user-generated content. Users can use songs they make via Udio for a commercial purpose, “as long as the content does not contain copyrighted material that you do not own or have explicit permission to use and as long as you properly indicate that the content was generated using Udio”.

Owning the commercial rights to songs made by Udio and Suno could provide users with some incentive to pay for a monthly subscription. However, courts in the US and the EU have so far refused to grant copyright protection for AI-generated artwork, and accordingly, songs generated with prompts in Suno and Udio could likely be legally monetized and licensed by anyone as the legal landscape looks today.  

AI Music Is Not Art