AI tracks a legal labyrinth
At first it really does sound like Frank Sinatra. “It’s Christmas time,” croons Ol’ Blue Eyes in his unmistakable baritone, “and you know what that means.”
So far, so good. Then things start to get weird. “Ohhh, the catch-up time,” he sings, as the strings in the background take on a horror-movie squeal. “As I light the tree to shield you . . .” The melody comes apart, fragmenting into echoing yelps of trumpet and sax.
Soon Sinatra himself is cracking and gargling, at one point sounding as if he has just unhooked his jaw like a snake. By the time he descends into nightmarish scat singing, punctuated by the disconcerting refrain “it’s hot tub time”, one thing is clear: whatever has stolen his voice is not human.
All that may sound like a distinct failure for the world’s most advanced music-generating artificial intelligence (AI). But this system, known as Jukebox, has already sent delighted ripples through the world of AI development since it was unveiled this month, having made rapid progress since its inception last August.
AI has been writing music since 1957, when a university computer the size of a VW bus spat out the Illiac Suite for String Quartet. There is now even an AI song contest, organised by Dutch broadcaster VPRO and being run this week, where computer scientists pit their algorithms against each other to find out which can compose the best tune.
However, no system has ever achieved Jukebox’s uncanny ability to capture, even partially, the sound of specific artists and specific genres.
“The moment you sent me the link, I’ve been playing [these songs] non-stop,” says Peter Alau, once a classical oboist in training and now director of business development at the Londonbased deep learning firm Spirit AI. “They are fascinating — some of them are just dead-on, and other ones are God-awful.”
Tell Jukebox to imitate David Bowie or Katy Perry and it just about gets them right — as long as they don’t need to form words. Tell it to cover Ed Sheeran’s The Shape of You in the voice of Simon and Garfunkel and it does sound like them — if they got drunk and didn’t bother listening to the original. It can even manage the verbal complexity of hip-hop, doing a passable impression of Kanye West rapping Eminem’s Lose Yourself.
Piero Scaruffi, a Silicon Valley software designer and prolific pop historian, is not impressed. “All of them sound like the work of an amateur who is using a laptop to make music in the style of famous people. I suspect there are many such amateurs, especially in high schools,” he says.
“The Beatles song sounds like a collage of snippets from a variety of their songs and then remixed in a lowbudget studio . . . I wouldn’t recognise it as a Beatles song, and not even as a remix of a Beatles song.”
Yet if Jukebox and similar technologies keep improving, they will raise difficult philosophical and ethical questions, as well as legal and economic ones. Though an imperfectly reanimated Elvis might sound creepy now, future AI might be able to convincingly generate new songs by long-dead artists. “It’s more than technically plausible; I think it’s inevitable,” says Alau. “There will come a time when whoever owns the rights to an album will say there is more money to be made and there is another album that could be released. “Taken to its logical conclusion, it will make Hollywood movie licensing look like child’s play. It’s going to be extensive and so convoluted that the legal community is going to have to create a whole new branch of law.”
Of course imitating existing music is very different to creating wholly new and fresh music, and Alau does not believe AI will ever manage the latter within his lifetime.
A neural network might be able to replicate old Beatles albums, but it cannot imagine what the Beatles would have done if they had kept playing, adapting and changing.