The era of AI music is coming, a few are happy and a few are sad

At first glance, "Simplicity's Delight" sounds like a catchy commercial pop song for Velveeta cheese. The singer sings the praises of "a plain piece of cheese" to light guitars and upbeat percussion. But if you look closely, you'll see that the sound is a bit like a computer-synthesized, but it's also possible that it's just Autotune. What really reveals the singer's fear of being inhuman is the second chorus, in which the word "tangy" is mispronounced.

There is now a new group of AI startups focused on building automated music creation tools, and the song above was written by software from one of them, Suno Inc. With a short text command, Suno can produce real-life vocal music in seconds, from dreamy electro-pop about a romantic breakup to an unplugged ditty that celebrates delicious fermented dairy products. It is reported that a complete AI album suspected to be produced by Suno has appeared on music platforms such as Spotify.

Generative AI is already being used to produce a wide range of content, including text, images, and videos. Music is becoming the next object, and everyone can be expected to express themselves through songs. But AI startups have also heightened concerns about musicians' livelihoods, and soon, the internet could be flooded with half-baked computer-synthesized songs. It will also test the tolerance of record labels, which have long been reluctant to take the tech industry to court.

Musicians and record labels have seen AI as a potential existential threat. In April, hundreds of musicians and bands, including Billie Eilish, Miranda Lambert and Aerosmith, signed an open letter through the nonprofit Artist Rights Alliance, urging AI developers, tech companies and others to stop using AI to 'infringe on and disparage musicians' rights.

Recently, Universal Music Group (UMG) reached an agreement with TikTok to strengthen its protection against AI music, after the company withdrew its songs from the platform due to concerns that TikTok was "flooded with AI-generated music." Last year, Universal condemned a hit song that used AI to replicate the style of Drake and The Weeknd. Dozens of music publishers have also previously sued OpenAI's rival Anthropic, accusing its chatbot of scraping lyrics without permission.

In recent years, leading AI companies such as OpenAI and Alphabet's Google have demonstrated the ability of AI to produce music, but so far, they have yet to launch a consumer-facing product. Google's DeepMind, for example, unveiled a music creation tool called Lyria last November, but has yet to release it to the public. The company said it was "engaged" with musicians and the music industry to advance "no-responsibility development".

In this context, the current AI music revolution is largely led by smaller companies. Founded in 2022 and headquartered in Cambridge, Massachusetts, Suno released its music production software last year, with a new version recently released that allows users to produce songs up to two minutes long. Similarly, Udio, founded by former Google DeepMind researchers and engineers, launched a beta version of its software last month that can output about 30 seconds of music clips.

"Two or three years ago, the technology was immature," said David Ding, co-founder and CEO of Udio, who has worked on DeepMind's generative AI projects, including Lyria. ”

Udio is backed by big names in technology and music. The company raised a $10 million seed round led by Andreessen Horowitz with participation from Instagram co-founder Mike Krieger, musicians will.i.am and Common, music producer Tay Keith and publisher UnitedMasters. Suno did not disclose details of the financing.

In order to attract users, both companies are currently offering free services: in the beta phase, users of Udio can generate 1,200 songs per month for free, and users of Suno can generate 10 songs per day, and become a paid member to get more rights to generate more songs. Both companies execute each user instruction twice to produce a different version.

Udio co-founder Andrew Sanchez said that in the first two weeks of the software's launch, more than 600,000 people tried it, and users generated an average of 10 songs per second. (No comparable data was provided by Suno.) Udio's software is also iterating at a rapid pace: several new features have been introduced in the past few weeks, including the ability to extend songs to 15 minutes.

Keenan Freyberg, co-founder of Suno, says that people often use Suno to compose a song for a friend or family member (or together) before exploring other features of the software. Teachers used Suno to create songs to supplement their lessons, and Palantir Technologies Inc., a materials software company, used it to create a country music for a recent software bootcamp.

Songs can be catchy, quirky, or both. A song titled "Rat-Contraceptive Disco" satirizes New York Times reports about New York City's efforts to stop rats from breeding with birth control. Excerpt from the lyrics: "Forget the rat poison, traps are useless." Why don't you have a disco revolution and let the gadgets appear. ”

"We felt like we'd be walking on thin ice because we couldn't predict how people would use these tools." Freiberg said.

But as AI has become more embedded in the creative realm, the technology has begun to clash with the entertainment industry and its copyright lawyers. Companies such as Midjourney, OpenAI, and Stability AI use datasets that include images captured from the web when building AI models that produce multimedia content. They argued that the practice was protected by the fair use doctrine of U.S. copyright law, but it still sparked outrage and lawsuits.

Pamela Samuelson, a digital rights expert and law professor at the University of California, Berkeley, said generative AI companies have a legitimate use defense for using works as training data. She argued, however, that the way the court viewed music might be different from works such as calculators, program codes, words or images. "The type of data can be important," Samuelson said, "and the court may make a distinction based on that." ”

Neither Suno nor Udio disclosed specific training data for their AI systems. Ding said Udio uses publicly available data from the Internet. Mikey Shulman, Suno's co-founder, said that in a way, the company believes that training data is even more important than how to build AI software, "so we keep it strictly confidential." But Schulman said Suno's approach was "legal" and "largely consistent with the lines of other companies"

Keenan Freyberg, co-founder of Suno, said, "We feel like we're walking on thin ice because we can't predict how people will use these tools."

This secrecy has upset Ed Newton-Rex, the chief executive of the nonprofit Fairly Trained, which provides certification for AI models trained with licensed data. Newton Rex, who was in charge of Stability AI's music production products, found it easy to use software from both companies to produce a large number of works that closely resembled copyrighted music. For example, he is able to produce songs that resemble bands and musicians such as Queen, Abba, Oasis, Blink-182 and Ed Sheeran.

"We don't know what their training data is, but if it's copyrighted work and they're building a product that competes with those works, it's hard to believe that they really respect musicians," Newton-Rex said. Sanchez said his company is in dialogue with industry stakeholders, including musicians and rights holders, "to ensure that our technology benefits all musicians and creators." One record label official, who asked not to be named, said they were open to a deal with a responsible AI company.

Suno declined to comment on talks with the music industry, but Schulman said the company was considering how to compensate musicians and was closely monitoring the changing legal landscape. "We really want to deal with this in a way that is fair to everyone," he said.

For this reason, Suno currently prohibits the use of creative instructions that include the names of artists, and Udio replaces them with other descriptors. For example, when a user asks for a "sentimental pop song in the style of Billie Eilish" about the difficulty of choosing a yogurt flavor at a grocery store, Udio replaces the musician's name with adjectives such as "folk pop" and "indie pop." It also includes a blue "Artist replaced" label in the track content.

Despite the concerns of musicians, the companies point out that it won't be long before a large group of people will be able to create professional-sounding music using at least some of the AI tools. "In the past, many people may have been excluded from music creation for reasons such as financial," Sanchez said, "and we thought this tool would remove their barriers." ’