Testing Spotify's AI playlisting feature

So, I've been making playlists for 10+ years now, probably edging towards 15. I even wrote a blog post called The Power of Playlisting when I was 16. It's no Shakespeare, but it sums up my feelings towards playlists pretty well. They're not just lists, they're moments in time and a very powerful tool for recalling memories.

I like that I personally built them when I was experiencing exam stress, a summer holiday, anything. I particularly relish the playlist I made to accompany my first solo drive after I passed my test. A year or two ago, I began the quest of building 'a chronology', a playlist that comprises of every song that has had a significant impact on me in chronological order - with some guessing involved. Yeah.

So I'm a playlist nerd. What's the hook?

Daniel Ek, the CEO of Spotify, has announced a new AI Playlisting tool. The vision is Chat GPT for playlists - you type in a prompt for the vibe you want, and it makes a 30-song playlist for you. Handy, right? Yes, but perhaps a little against my playlist manifesto. It's currently in its Beta phase in the UK and AU markets, so I felt like I was a brilliant candidate to give it a try.

I imagine the typical perceived use case is;

1. Person A is having a dinner party with their parents-in-law/colleagues/son's violin teacher, and is keen to impress.

2. Person A wants to play dinner appropriate music, but is secretly a raging emo.

3. Person A uses AI tool, types in 'chill dinner music', and gets a ton of Snow Patrol, Leon Bridges and David Gray placed into a handy 30-song playlist that lasts a couple of hours.

4. Said guests think this person is very grown-up indeed.

And crumbs, you didn't get the right vibe you wanted? Well, not to worry - you can follow up with more suggestions, à la Chat GPT, to curate the ideal flavour.

Seems cool right? Okay, let's give it a try...

NB: This feature is in Beta, so I'm testing it mostly for fun and no criticism is meant wholeheartedly! It's still far more clever than me and I love our Spotify overlords.

User Case 1: Situational

In this user case, I asked for 'chill and sensible music for when the parents in law come over for dinner'. I got suggested a ton of my usual listening habits - fair enough, it wants to keep it reasonably to your taste. However, it was stuff like Mac Demarco's 'Cooking Up Something Good', which is about his father's methamphetamine habit. It doesn't scream conservative - and somewhat surprisingly, I actually don't listen to that many songs about drugs - so it seemed random to pick out.

With the amount of very wholesome folk (First Aid Kit), middle of the road pop (Beck) or just classics (Bowie, Beatles) in my listening repertoire - I just feel like there were more options. Spotify, I appreciate you missing out the anti-disestablishment Radiohead tracks, but you're allowed to push the boundaries.

I asked Spotify to make it 'more boring please, they are conservative', to which computer said no.

I asked again, in simpler terms: 'more boring, and less quirky'.

This time, it fed me a ton of instrumental tracks that I listen to already - Khruangbin, Kishi Bashi, Intro by the xx and some Aphex Twin. I wouldn't say Aphex Twin screams 'chill' most the time, but it was Avril 14th, so I'll allow it. So perhaps, to be conservative, you need no lyrics at all, in my case of taste - or according to Spotify's all-knowing AI.

Hilariously, the AI playlist name was 'Conservative Dinner Tunes'. In my mind, David Cameron's annual hog roast probably is filled with 'Jump for my Love' and the like, but my in-laws are getting a piano cover of Where is my Mind.

So, thus far, it appears that Spotify is mainly keeping aligned with your habitual listening and 'liked songs', which sort of defeats the purpose for someone looking for something beyond their usual tastes.

But this is only one user case, so...onward?

User Case 2: Genres

Right, so it appears that Spotify does tend to use your tastes to inform the output playlist. Let's test that a little by asking it to create genre-specific playlists. Here are two tests:

1. 'Jazz' - a genre I don't really engage with much, if at all, with Spotify. Will it use tenuous links to match my tastes?

2. 'Bubblegum pop' - I engage with a lot of pop music, but not in the very generic 'bubblegum' sense - more indie and alt pop. Love me some TayTay though.

Results:

Spotify AI is aware I don't engage with jazz, and hasn't delivered things that 'could, at a push, be considered jazz' that align with my tastes. Instead, it's given me a list of songs that are reasonably generic and recognisable to a non-jazz person, but are still definitely proper jazz. Nice one, tick.

'Bubblegum pop' is a bit more confusing. I think the intention here is that the machine tried to create a blend of tracks I should recognise, alongside some more classical bubblegum pop tracks.

However, the two songs which I haven't ever engaged with on Spotify, ie. Green Tamborine and I Think I Love You are not Bubblegum Pop at all, at least not to me. However, according to Google, Green Tamborine is 'credited as being the first bubblegum pop chart-topper'. Interesting! It's a novelty song, as is I Think I Love You. Perhaps that's the parameter here. So there's some Google searching at play.

However, Ashnikko (absolutely hyper-pop, pop-rap) and Janelle Monae (R&B), both songs I listen to and have been inserted as my 'familiar faces' - don't seem to align with Bubblegum Pop at all. The other 'familiar faces' (SOPHIE, Dua Lipa, MUNA) don't explicitly fit the genre - but if a gun was put to my head and asked to refer to them as 'Bubblegum Pop', I wouldn't push back too much. Taylor Swift - we're there. Half marks on this one, Spotify.

User Case 3: 'Complex' terms and jargon

Okay. I'm big on my quirky music terms. 2016 B-Town (the Birmingham indie scene), Landfill Indie, Queer Pop. I spent a long time looking up to other music journalists and the terms they were coining. I want to know how much Spotify actually recognises these terms, especially as there are editorial playlists named 'Slow Sadcore Mix', 'Drift Phonk' and the like.

Let's keep it simple and ask for Indietronica. This isn't just one simple, cookie cutter genre, but a very well-known term amongst most music fans, and not a challenging one to translate for someone unfamiliar. My belief was that if it didn't come back with some Hot Chip, the AI is not really doing its job. And it didn't! See left.

I wouldn't call early Bombay Bicycle Club or The xx Indietronica. LCD Soundsystem for sure, Vampire Weekend I can lightly agree, but Rebellion by Arcade Fire? That song is loaded with pure orchestral elements - not a synth in sight!

Interestingly, the AI messaging fed it back to me as 'noughties indietronica' and didn't seem to notice that it was a blend of two phrases. I think it far more picked up on the 'indie' and 'noughties' and went from there.

The machine learning is working overtime though. I asked again an hour later, and it came back with far more of an indietronica vibe. Metronomy! MGMT! Empire of the Sun! And, praise the lord, Hot Chip!

I omitted the 'noughties' this time, hence the newer 2010s tracks - Broken Bells, Glass Animals and recent Bombay.

The AI also acknowledges in its response that this is 'blending indie and electronic vibes', meaning that it has understood the prompt and the genre. It's fair to say that MGMT, Metronomy and Hot Chip and considered staples of this genre, so I think the machine learning has worked better in this case.

Cool, one phrase learnt!

Shall we step it up a notch?

I'm going to ask for 'Landfill Indie'. This term, coined by Andrew Harrison at Word Magazine in the mid-noughties, is admittedly very snobbish. It's now used to group the huge number of rather generic guitar bands that emerged in the 00s. It's also something a malicious 40-something ageing music journo, embittered by his lack of success in the performance world, would call trendy bands following a pattern of upbeat guitar music. Bands that are perhaps a bit formulaic, and might get lost in the riffraff of noughties indie - but are not objectively bad. Anyway, what does Spotify AI make of phrases loaded with opinion?

First prompt:' Indie Landfill of the Noughties'

Response: 'Dive into the indie landfill of the noughties with this nostalgic playlist...'

Second prompt: 'More one hit wonders, more generic'

Response: 'I got it wrong this time. Making a fresh playlist...

Here's a playlist with indie one-hit wonders from the noughties to satisfy your request.'

Both outcomes below:

I understand this is all down to opinion. Whilst Kings of Leon were certainly not classified as landfill, I could see a disgruntled journalist labelling some of the others so - namely Stereophonics and Kasabian. The second iteration is very similar, and includes The Kooks, who were definitely, regardless of whether it was fair, referred to as 'landfill indie'.

Moreover, I wouldn't call these one-hit wonders. Float On perhaps, but MGMT had several hits, as did KoL. Also, they've included TWO Arcade Fire tracks, which kind of goes against the point.

However, whilst this is 'noughties indie', they haven't chosen to include the big dogs - Arctic Monkeys, The Strokes, The Libertines - the gatekeepers of 2000s Indie. This playlist is very XFM-ified, without all the frantic obsession with Oasis. So there's definitely been a choice around these being 'smaller hits'.

I would have imagined that the AI would use some context from Google to determine the meaning of 'landfill indie'; perhaps cross-referencing the classic online music mags, pulling something from a Vice article like this one. I'm not sure what method the AI is using here, but it's certainly not that. It seems like it's still trying to keep in with my tastes, as there are several tracks on here which are in my frequent rotation.

Conclusion

I think we've found a few trends here. It's clear to me that this is in Beta mode as it doesn't seem able to completely fulfil my proposed user cases as of yet. But some key points from my experience:

1. For most prompts, this AI will still use your existing personalised algorithm, formed around your tastes and previous listening, to inform around 50-75% of the tracks it picks, unless it's a genre you don't listen to. This means that it is less useful for some more specific user cases.

2. There is not much nuance when it comes to subgenres.

3. It still needs to learn a lot more specialised terms from particular subcultures and regions, but this will hopefully come as more users begin experimenting with it.

4. It seems to be learning, rather quickly - trying similar prompts within an hour got vastly different results.

I hope this was as interesting to someone to read, as I had writing it. If I experiment any more, I'll be sure to add onto this post.

Happy Monday! El x