AI has come for music. So far, there’s a Great Big Sea between people and algorithms

ST. JOHN’S, N.L. — The production is slick, the drums are on point, and the vocals sound great, but a titan of Newfoundland and Labrador’s music scene hears something off about “It Could Be Worse” and “Tales of The Atlantic,” two songs generated in less than a minute by a powerful algorithm.

“It’s a country singer, so that’s wrong. And the lyrics don’t really rhyme,” said Bob Hallett, a founding member of Newfoundland folk-rock band Great Big Sea. “It just sounds sort of strange.”

Hallett had just finished listening to the rollicking tunes, which were created using a generative artificial intelligence tool called Suno using prompts that could describe any Great Big Sea jam: Celtic, folk, lively, passionate.

They didn’t hit the mark, Hallett said. On a scale of one to 10, with 10 being the band’s hit “Ordinary Day,” he gave the songs a two.

But experts say that rating could increase quickly. Technology like Suno’s is advancing swiftly, and its output is only getting better, said Jimmy Lin, a professor and director at the University of Waterloo’s artificial intelligence institute. 

Suno is one of several companies building generative artificial intelligence software that allows users to create original songs using text prompts. People can create instrumental tracks or songs with lyrics, which can be generated by the program or supplied by the user. But if the user offers copyrighted lyrics — the first lines of Nirvana’s “Smells Like Teen Spirit,” for example — the program won’t make the song.

Suno also doesn’t make songs intended to sound like other artists’ work, such as “Heart on My Sleeve,” the AI-generatedsong using unauthorized sound-alike vocals of Drake and the Weeknd which caused controversy in the music industry last year. When asked to make “a Great Big Sea song about cod fishing,” the resulting Suno tunes had a pensive, Celtic flair, but they sounded nothing like an authentic track by the band.

Google is working on similar software, called MusicFX, which can be sampled through its AI Test Kitchen site. And Adobe unveiled Project Music GenAI Control last month, which it described as an “early-stage generative AI music generation and editing tool.”

In December, Microsoft introduced a Suno-powered song generator for its Copilot chatbot, which is a program that uses artificial intelligence to simulate conversation with users.

The technology behind these programs is similar to that powering the ChatGPT chatbot. 

Lin said such programs use massive data sets to “train” algorithms, or step-by-step processes, to take any starting point and predict what the next one will be. So while chatbots trained with text can predict the next word in a written answer, a music-generating program is trained using sound to predict the next “acoustic sequence,” he said.

The New York Times sued Microsoft and ChatGPT owner OpenAI in December for using its stories to train programs.

Lin said companies behind AI platforms that make music could find themselves in similar trouble if they’ve trained their algorithms with work by artists who haven’t given consent or received compensation for their music to be used that way.

“It is an unresolved question whether this is fair use or not,” he said in an interview. “It’ll sort itself out. The court always does.”

Suno’s website does not indicate what data it used to build its program, and the company did not return a request for comment.

Hallett said he wouldn’t be surprised if the algorithms had learned from a few Newfoundland bands. He said the songs it produced had a few signature sonic markings, including tight melodies and heavy strumming on acoustic guitars, that he and his fellow producers have cultivated over the years recording albums for bands like the Ennis Sisters, Shanneyganock and The Fables.

He was unfazed, however, about the music being used to train these programs, noting that artists have long been contending with platforms like YouTube and Spotify that already cut deeply into musicians’ earnings.

“There’s a bit of a feeling of surrender about it all … it is so hard to police it,” Hallett said. “Creative work is really about driving your concert sales, or finding commercial placements. Even at the highest levels, people aren’t really making money anymore selling records.”

Lin said AI-generated songs will likely be used by advertisers who need a catchy jingle for a commercial. And he believes that could start happening soon given the dizzying pace at which these tools are evolving.

“We’re not talking about years or decades. We’re talking about months,” Lin said.

But Hallett said anyone looking to use music to connect with an audience would be best served by human beings.  

“It’s easy to become scared of AI,” he said. “But all of us are drawn to sincerity in music. We want to hear people who are telling us a real story and delivering real emotions. And the computer just can’t do that.”

 This report by The Canadian Press was first published March 3, 2024.

— With files from The Associated Press

Sarah Smellie, The Canadian Press