During the pandemic, I have taken to working more with headphones, often exploring the many musical genres offered by YouTube. Having established my pre-existing liking for guitars and 80s synths, its ‘algorithm’ led me down the path to ‘post-punk,’ Serbian ‘drum-and-jazz,’ suggested I try Norwegian ‘darkwave,’ and presented me with ‘Zamrock,’ a mixture of Western rock and traditional music from 1970s Zambia. Other algorithm trips took me to early electronic music, strange time-signatures and bands with incredible range. Much of this music would have been completely unaccessible to most listeners 20 years ago, and I’m extremely glad it’s out there, waiting to be discovered anew. Many others are too, with a common comment stating that ‘the algorithm has found a gem.’
Recently, the algorithm made a new suggestion: music written by Artificial Intelligence (AI). Not worrying too much about whether this had been its goal all along, I immediately clicked on ‘Smother.’1 This seemed to be a ‘new’ tune in the style of Nirvana, the 1990s grunge act best known for Smells Like Teen Spirit. ‘Smother’ was just like Nirvana, very catchy and well produced. It seemed impossible that it could have been written by a machine. Indeed, it was too good to be true. When I read the description, I saw that only the lyrics were ‘written’ by AI. They were the worst element of the song, an incoherent mash of the words that had been fed to the AI. The instruments, vocals and mixing were all the work of one talented human.
Disappointed that the computer had not done more, I looked for other examples and found ‘Daddy’s Car.’ 2 This is a tune in the style of the Beatles, arranged by researchers from Sony CSL. The lyrics, chord progression and melodies were all suggested by its Flow Machines AI system, with selection, arrangement and performance by humans, acting almost like the producer. The result is definitely a song, although it feels like the researchers have been forced to limit some of the AI’s wilder urges. Fed with 45 Beatles tracks, it is not surprising that there are references to the ‘taxman’ and ‘diamond skies.’ It seems that ‘daddy’ has been exploring the YouTube Algorithm too: “In daddy’s car, the sound’s so good, like something new, it turns me on.”
The same group also produced Mr Shadow,3 which has a computer-generated voice cut over a string and guitar arrangement. Its AI influences become clearer as it plays, prompting the comment: It sounds like the credits sequence for a film where the robots won.
But what if we could remove humans from the creative process? There are already AI symphony generators capable of generating original content, albeit within fairly strict parameters. Critics point out that due to the imposition of these rules, such pieces are generic and simply average out the feed material.4 They play it too safe to offer creative value.
There are also examples of systems like OpenAI Jukebox system ‘continuing’ well-known songs after ‘hearing’ a sample. These demonstrate an unchained AI basically making stuff up and, in many cases, going off the rails. Most sound like a cover band that has only listened to the song once before coming on stage, if that. The Don’t Stop Me Now (Queen) extensions are some of the more amusing.5 It is a remarkable feat of coding that the output sounds like music at all. Indeed, there are even flashes of ‘inspiration,’ where you hear a chord and melody combination that could be the basis of a real song. In the right hands, such approaches could act as an artificial writing partner, muse, or a way around writers’ block.
Right now though, whether an AI-generated song is ‘bad’ or ‘good’ is up to the listener. Here, there are many opportunities for disagreement. There is no statistically correct move in music like there is in playing games like Chess or Go, or in most other processes taken on by AI. Even something as complex as running a cement plant offers clear targets that can inform the system whether or not it is on the right track.
Music and other creative pursuits do not offer this kind of numerical feedback, although there may be sufficient advancements that make it possible for AI to judge its own creations in the future. Just look at the algorithm that brought me to the topic of AI music in the first place: It knew what I wanted to hear! In the future, AI could even generate music for single-use listening. Just ask for Latin-infused jazz versions of simulated Led Zeppelin songs and the algorithm will oblige. Would that be original and creative, or just random guesses? And, more importantly, would it be worth listening to?
1. https://www.youtube.com/watch?v=GogY7RQFFus
2. https://www.youtube.com/watch?v=LSHZ_b05W7o
3. https://www.youtube.com/watch?v=lcGYEXJqun8
4. https://www.youtube.com/watch?v=Mxa6k3AgNqs
5. https://www.youtube.com/watch?v=mDH2KD8V5Ys