As music streaming has quickly become ubiquitous, and more readily connected to our iPhones and Sonos speakers than ever, we're relying on AI to suggest new music for us to listen to. Joe Sparrow asks if we're actually hoping for something more...

Cast the first stone

Music is different to other artforms. It can run in parallel to other activities. You can listen to music whilst doing other things - working, running, idly plucking at nasal hair - and still enjoy it to the fullest extent.

You might be re-grouting the bathroom, yet your heart can still soar at a piece of music while you’re doing it. Now, try appreciating a Tracy Emin installation on your hands and knees in the shower. Music is just different like that.

And you don't need to explain musical favourites: they just are. You should never need to explain why, to pick an example entirely at random, the debut LP by mid-90s cagoule-wearing Britpop also-rans Cast is your go-to hungover-on-a-Saturday-morning album, even though you find yourself repeatedly doing exactly that.

Harder Better Faster Stronger

So if we have firm favourites, why do we need to discover new music? Because we’re dumb, restless humans who can never get enough of anything. Also, how did you find that favourite in the first place, smart-arse?

In the past it was simple: pick a radio show, listen every day, and buy the songs you like the most. Or get a recommendation from a trusted friend: the gold standard of recommendations, almost guaranteeing a listen.

Today it’s more complicated. More and more often, we find new music when your streaming service of choice nudges us towards new songs and artists. We're happily allowing machine learning to introduce new art into our lives. And it's frankly very good at it. But what if it's not what we really want?

99 Problems?

The myriad music discovery systems are all essentially there to solve the single problem of finding the unknown-unknowns. When we say something like “I want to hear something new,” what we’re actually saying is: “I want to hear something that I will definitely like that I don’t yet know I like, or expect, or know already.” It’s a big ask.

For instance: how readily will you commit to a song like The Killing Moon - the “greatest song ever written” according to the humble Ian McCulloch, the man who wrote it - if you don’t know anything about 80s UK rock music, let alone the absurdist mystique and legend that swirls about Echo and the Bunnymen like bong-smoke?

Incidentally, Ian McCulloch is right - it is the best song - but you need to arrive at that conclusion yourself. It’s this cavernous gap between opinion and fact that music-suggestion tech hopes to solve.

The vital connection is made

Music discovery is a bit different to discovering other art in other media. TV shows can be suggested on a like-for-like basis: enjoyed Game Of Thrones? Here’s another quasi-historic guts-’n'-mud TV show with dragons you can use to while away another 100 hours of your life.

One can hop between movies on the strength of a star, or books on the strength of the author. But music seems to resist simple “if you like that, you’ll like this” suggestions. Maybe it’s because pop music jiggles to life nebulous feelings as much as simply reproducing a catchy tune. Half the art of recording music is to capture lightning in a bottle: pop music is about inducing powerful emotions at will with each repeat-play.

And because big, deep emotions are at stirred, the barrier of entry is high. How do you let a new song into your heart when there are just so many of them? In the world of pop, scepticism is high, and tolerance is low. There’s a reason the best pop songs reach the chorus in the first 30 seconds - we just won’t wait any longer.

Automated music discovery systems tend to take one of two approaches: trust the wisdom of the crowd, or attempt to know you better than you know yourself.

Services that take the former approach - like The Hype Machine or - gauge what others are listening to and suggest what you might want. They work well because there’s a tacit understanding that this method of discovery has a low hit-rate: users expect misses as well as discovering new songs they love. But that means there are built-in limitations to this approach too: most people just want the hits.

The latter system - aiming to know our own music taste better than we do - is aimed at this bigger, trickier audience. And these systems can be eerily accurate when it comes to picking the next song for you. It’s because their tendrils, analysing you and your music taste, extend further than you’d expect - into music and your listening behaviour.

Spotify’s Discover Weekly, launched in 2015, is notable both for how on-point its recommendations are and for the feeling that it has somehow figured you out, pulling songs that you’d forgotten you love from the crevices of your memory. Here's how that magic process was explained to Hackernoon:

”To create Discover Weekly, there are three main types of recommendation models that Spotify employs:
• Collaborative Filtering models (i.e. the ones that originally used), which work by analyzing your behavior and others’ behavior.
• Natural Language Processing (NLP) models, which work by analyzing text
• Audio models, which work by analyzing the raw audio tracks themselves.”

It sounds (and is) complex, but Spotify uses a few simple ideas (which result in complex code, eye-watering algorithms and some mind-boggling shuffling of datasets) to spew out personalised recommendations.

Recently, Spotify launched Your Time Capsule - with the vague tagline “a personalised playlist to take you back in time.” Stirring nostalgia - that most human of emotions - is a scientifically-proven smart move: the songs we liked in the past have a much stronger emotional pull over us than equivalently "good" songs from today.

I found that it pulled together a pretty good “personal best-of” playlist based on songs I like, but not anchored at a particular point in my life. On the other hand, friends told me how weird it felt to be dragged back to a time and place by those eerie algorithmic fingers.

Consider too, one of today's subtlest and most effective forms of recommendation: autoplay, when the music silently clicks over from something you chose to something you didn’t. How many times have suddenly realised that the last hour’s music was not picked by you? Was it when a great, new, song was played?

And there's more to come: Google Music says it uses computational AI to fine-tune its recommendations based on all that delicious data they have gathered on you over the years. Searching the web for "music suggestions" reveals a glut of different approaches, all with varying degrees of success.

Of course, one big reason that streaming platforms want you to listen more is because they want you to keep subscribing to the service (and it’s effective - back in 2015, Discover Weekly created 1.7 billion streams in under six months.)

It feels like the tech giants at least have almost got this licked, so far as they are using their bespoke algorithms to create slightly different flavours of recommendation sausage. So what if we want something else to eat?

Radio Head

Or maybe in an era of automated-this and algorithmic-that, we’ve come full-circle. Two years ago Apple Music launched - improbably - a radio station, Beats 1, which they claim is now “the biggest in the world”. In this model of music discovery, humans - expensive, narrow-band algorithms with interesting voices - are important again.

And it works: Beats 1 is pretty good, although who knows how many people actually listen to “It’s Electric! - a weekly window into the mind of Metallica’s Lars Ulrich.”

Amid all this is the odd resurgence of vinyl, a very tangible form of music discovery in an ephemeral age. And while the vinyl sales charts suggest the vinyl-buying demographic is your dad, it’s equally likely that vinyl is a useful form of discovery, as long as the records being discovered are the same ones their parents used to play.

Both of these throwbacks suggest that, amidst the waves of auto-suggestion, listeners are hankering after something a bit more... human.

And why not - you don't wonder if your Discover Weekly playlist is flirting with you when it suggests you should listen to the 12" Club Mix of Take Me I'm Yours, but you might if Steve from HR did.

Maybe today, that “trusted friend” is anyone and everyone, just like how everyone is your friend and/or enemy online, tantalisingly within kissing/punching distance.

On one hand, we are all “musical discovery influencers” now: on a server farm somewhere, my data is meshing with yours and creating baby playlists which will open someone's eyes to music they never knew would move their soul (or introducing an unfathomable amount of late-90s UK garage hits to their playlist.)

And yet it feels like something real is missing too.

If music really is a different type of art, perhaps that introductory moment needs anchoring differently too. There's a balance to be struck between AI and DJ. Meanwhile, game the system and get the best of both worlds: re-purpose a sultry playlist as one of your own and send it to your office crush. Or Steve from HR.