Music algorithms failed us
Or how streaming promised so much, but delivered so little
Of all the technological innovations that fell flat on their stupid faces, music recommendation algorithms may have fallen the flattest and stupidest. I’ll even go one step further: they’ve been detrimental to the general enjoyment and discovery of music.
To get all flat-cap-and-pipe-and-whippet-and-long-days-down-the-mine about it, the main way to find new music The Olden Days was fairly simple: family, friends, and media — normally in that order.
It’s how I began. Some of my earliest memories are music related, dancing and racing around the house to Queen’s ‘Headlong’ or whatever other tunes my parents spun. From there, it was those playground ditties, tracks that people share at school or you hear on the radio. You know, like Robbie Williams’ Rock DJ.’
At this point, music obsessives then tend to start treading their own path, veering away from the tastes of family and friends. In my time, this meant engaging with radio, television, magazines, and, eventually, the big daddy of them all, the internet.
This played out like a sliding scale. While TV introduced me to some new music, the majority of channels played a pretty mainstream selection of tunes. Even something like Kerrang! (a rock and metal channel) broadly struck to the big hitters.
Radio could be more obscure. While daytime Radio 1 was geared towards mainstream sounds, a lot of interesting and underground music emerged in the wee hours. I’d schedule my stereo to record late-night John Peel or Mary Anne Hobbs shows and then comb through those for something I liked, something that spoke to me.
Arguably, though, it was magazines that made the biggest impact on my music discovery path. I subscribed a brought so many, from NME to Kerrang!, from Guitar World to The Stool Pigeon. I read them from cover-to-cover, hoovering up every scrap of information I could. There was no easy way to actually listen to much of the music they discussed, so I read a description of an artist that sounded cool, I’d go and buy the CD. Sometimes, I’d go into a music a store and pick up an album because I liked the cover.
All of this feels strange to think about now because the internet utterly upended how I discovered new music.
Most magazines and radio shows were tied to genre. It was an indie channel, or rock publication, meaning it had clear borders on what it did or didn’t write about. The internet didn’t subscribe to this notion.
Through niche publications, forums, and private trackers I was suddenly introduced to thousands of people with unique and idiosyncratic music tastes. It was like the part in The Wizard of Oz when the world blooms into colour; I heard obscure outlaw country; bootlegged records from 70s Ghanaian artists; forgotten naive rock classics — things that were totally outside my point of reference before being subjected to the glorious opinions of other stinky music nerds online.
Things only got better when streaming services came along and (metaphorically) wheeled a shopping trolley of fireworks into the mix.
Getting into music at the start of the online era had a frizzing dissonance at its core; although you could talk about any track, it wasn’t a just few keyboard clacks away. To get a song, you had to go treasure hunting, boot up Limewire, search a torrent site, or beg strangers. Combine this with slow internet, viruses, and janky software, it meant there wasn’t often a clear route to just listening to a song someone said you just had to hear.
Spotify’s wet, hot birth changed this. It ruled. And by many factors, it still does. Having the entirety of the world’s music at your fingertips remains wonderful and magical — it’s a true modern marvel.
For a moment, I was convinced that This Was It. Music discovery was solved, the taste of the general public would skyrocket. Even Spotify seemed in on the game. Not only were interesting user-created playlists highlighted, the company actually hired curation teams to create their own.
The vibes, though, couldn’t last. Nothing good online can — apart from Wikipedia. Praise Wikipedia.
Lo and behold, nowadays, Spotify’s pretty awful for finding new music (for some rather seismic reasons). Sure, there’s still a user-driven playlist scene, but the platform itself is increasingly geared to slop pumping. The goal is to keep you listening, not encourage exploration.
The beginning of the end of Spotify being a place for finding music rather than searching for it coincided with the launch of Discovery Weekly, a tool ironically meant to promote new music to people. Released in 2015, it was billed be a personal DJ, something that could recommend tunes which fit your taste.
The problem is it’s shit and misunderstands what loving music actually means.
It all comes down to how algorithms reduce taste to data. This excellent piece goes into depth about how these systems work, but we’ll run through the main points.
Effectively, music recommendation algorithms operate on two main plains: content and collaborative filtering. Content filtering is about digging into the nature of a track, while collaborative filtering concentrates on how it fits into the wider ecosystem.
For example, content filtering relies on both using tags (like genre, record label, etc.) and analysis of the song itself, such as BPM or an algorithm-defined feature such “danceability,” something that can be ascertained by tracking things like rhythm.
If content filtering is focused on internal elements of the music, collaborative filtering is about the the external elements. This looks at how a track fits into a network of people and tastes. Ergo, if individuals with similar listening profiles both start repeatedly playing a new song, it’ll get tagged with that information. Someone else with that listening profile are then more likely to be recommended that song.
Let’s take a surface-level example. It’s likely that fans of The Smiths will also listen to Morrissey. This connects those two artists innately. So, if Spotify notices someone listening to a lot of The Smiths, but not Morrissey, Discovery Weekly will probably recommend them the singer. It’s more complex than that of course, but that’s the basic principle.
There’s a hulking elephant wrecking the room though: similarity. Recommendations are connection networks. To put that another way, streaming services recommend you music that sounds similar to other music you like, stuff that’s directly linked to your pre-existing taste. This creates a self-fulfilling prophecy where most of the music you’re shown is deeply similar to your pre-existing taste.
Think of it like an averaging machine. Spotify will show you what the majority of people with a similar music taste will listen to. It’s not going to show you something out there or interesting.
For example, if you listen to modern punk music (like Pup), it will predominately recommend you tracks from that genre. What it won’t do is suggest an outlaw blues musician like Leadbelly, someone who is spiritually connected to the ethos of punk, yet sounds nothing like Turnstile.
Part of this is the nature of these large scale models, and the other part of it is deliberate.
People want to listen to music that’s familiar to them, we’re biologically geared towards it. While, for a company like Spotify, keeping people listening and on the platform is paramount. It makes sense for them to “recommend” you music that’s similar to what you already like because it can blur into the background. But if something more left-field comes on? A track that sits outside your specific likes? There’s a strong chance you’ll tune in to what’s playing and turn it off.
And this is the crux: Spotify and Discover Weekly are no longer primed to help you find new music, instead they’re designed to keep you statically listening.
The best music I’ve been introduced often comes out of nowhere, unrelated on a surface level to other stuff I listen to. Instead, the connection has been more tenuous, more ethereal, more associated with a feeling or approach — not how fucking “danceable” it is.
I’m a firm believer that algorithmic recommendations can never be great. They work too literally, they’re too divorced from the idea of taste. They’re programs, averagers, networking machines, not beings of grace and style and substance like you and I.
Spotify’s biggest mistake was moving away from people, of relying on machines to recommend us art that explains how it feels to be alive.
We lost something. A wonderful future vanished. Spotify can still be an excellent tool for finding new music, but you have to work against it, not with it. And that’s a crying shame. Steaming could’ve been so much more, but, like man other things in the modern world, it reverted to a very specific goals: keeping us thoughtless, quiet, and consuming.