Mood Machine: The Rise of Spotify and the Costs of the Perfect Playlist

by Liz Pelly

I continue to use Spotify in the same way I used iTunes, and CDs and cassettes before them; I know what I want to hear, and then I listen to it. The difficulty for both me and Spotify, presumably, is figuring out what I want to hear. It’s the difference between following a treasure map—it has landmarks, a route, a clear goal—and finding the treasure map in the first place: that’s mostly kismet. There really aren’t maps to maps, most of the time, but once you have directions actually getting to the gold comes down to simply following the instructions.

Mood Machine: The Rise of Spotify and the Costs of the Perfect Playlist tells me that Spotify isn’t in the cartography business, and that listeners don’t want to follows maps or find treasure. My data would be dismissed as an outlier. Apparently, the main growth area for music streaming services is background noise, akin to the way I see the Baby Boomer generation flip on the TV the moment they arrive home simply to have something continuously spurt sound into the air. Nobody wants a map to real treasure—most are content with whatever pyrite can be shipped directly to their home—which is good for Spotify because they were pretty trash at constructing maps, anyway. And there is a finite amount of treasure out there in the world. It’s fair to be bad at procedurally generating treasure maps; honest treasure is both hard to find and, as the old saying goes, uniquely beheld by each person. How much worse, to put in the work to follow an intricate hidden path and find that someone buried a box of buttons, say? Treasure to some, trash to others, but when you’re in the treasure business, just hand people a reasonably glittery version of what they type into the search bar, and voilà, the ceaselessly expanding profits of platform capitalism! It seems much more challenging to large-scale point to actual treasure that exists somewhere out in the world, buried beyond the boundary of landscape your business controls. Easier, then, to just manufacture a theme park where you can keep popping open the same inoffensive chests filled with the same cheap tat for everyone.

If you’ve ever supplied the Spotify Algo with a Midjourney-esque prompt–“Rollerskating Fun Tunes for Children,” say–you’ll get back a bunch of trash. And letting the auto-select veer your listening off into the uncharted world will bring you back to port quite quickly. I vividly recall the exact moment when I shut down the ‘autoplay’ or ‘recommended music’ feature on my account: in the middle of an indistinguishable mush of soft, lady-with-guitar (my favorite genre) notes that were blobbing through the room, Capsize by FRENSHIP came slicing through the air. I thought to myself, “Hey, self, this song is great, but why is it here?” I knew I played it often via a prior playlist I had made, but having the “auto-play/discovery” model regurgitate a deep cut song I knew and liked and chose for myself a lot felt like pandering. So I shut down the auto-play feature—which rarely impacted me anyway because I tended to listen to a single album for a set amount of time and then close Spotify, hooray for me—and felt content in my 2009 decision to leave the Pandora “lean back” radio model for the “select your own songs” Spotify construct. Nowadays, if I want background music I have the ostensibly human-curated “chill JRPG classic nostalgia songs PS1 vol.2,” thank you youtube. 

In the process of checking the spelling for Capsize and FRENSHIP, I discovered that my story lacks a bit of the impact I intended–Capsize is not a niche song from my personal music history that slide itself into my discover mode: it has over 610 million streams on Spotify alone. Perhaps it illustrates a different point–the poor algos have to deal with human caprice and balance the fundamental misunderstandings of things human believe about the culture we consume. Even the most advanced machine-learned AGI simply cannot have the capacity to predict the small fallacies that most of us live by. No robot in the world is going to look at a song with a half-billion streams and assume a hypothetical playlist listener will react with surprise when they hear it served up. I legitimately shut down the autoplay because I thought I was getting served too-recursive, self-reinforced music choices when I heard one of the most popular songs I personally play regularly. You cannot make it make sense, and yet, here we are.

Other human-math incoherences that I pity algorhythms (sincerely, if they’re leaning in anyway, why don’t they call ‘discovery mode’ that???) for having to confront include the “shuffle” button: “random” settings need to be weighted, even on an old discman, because “clumping,” or repeated plays of the same song, wouldn’t feel random to a person. To a machine, though, that’s true random. Sorry, perfect code, you lose to bewildering human interpretation.

Another of my absolute favorite examples of math bending to the whims of humanity is the functional misinterpretations of percentages, specifically how they’ve been presented by videogames. Strategy RPGs have admitted to tipping the scales to match with (mostly young adult/juvenile) expectation and understanding of probability: the player does thousands of attacks over the course of a game, so if one were to miss hundreds of attacks it would feel bad. In some games, a display of a 90% hit rate is basically a “hit.”

Ike never misses

Imagine trying to explain the Monty Hall problem to a person raised on strategy game probability. Except X-com. X-com players always know to switch doors.

Now that the math portion of this examination is over, we can talk about music. Every time the sun comes up I’m in trouble, but every time I hear that Sharon Van Etten song, I’m delighted. Caprice. Does it and its odd rhythm fit into the offbeat pop-alt-folk user profile the computer has built for me? No, likely not. The algorhythm (it works) needs to take those 90% hit-rate shots on things that seem like they should fit, but no one can juice the outcomes when it comes to whim and fancy. Mood Machine: The Rise of Spotify and the Costs of the Perfect Playlist purports to uncover the math that makes Spotify run, but ends up walking itself out of the interesting-fact forest that dots the opening of the book—ghost composers filling playlists with stock music on the cheap; payola discovery modes lifting 30% from the pockets of new artists–and wandering back into the tilled and tended fields of mutual aid and platform capitalism. 

I am quick to remind myself regularly that every encounter with ideas is likely someone’s first, so kudos to this book for taking its mainstream pulpit (it was hard to get at the library, and harder still to keep) to cogently present on pressing social concerns, but I’ve heard this song before, and better. Mood Machine sings loudest when it details what is actually going on within the ones and zeroes of Spotify:

On the purely algorithmic front, new types of automated “Niche Mixes” overtook more screen space, but the music contained within them never seemed that different; “Sad Crying Mix,” “Sad Late Night Mix,” and “Lonely Sad Mix” were all different options, when I clicked around recently, but the offerings were mostly music from my listening history, with just a few new recommendations. The boxes within which Spotify segmented its recommendations were growing more niche and plentiful, but more than anything, it felt like an interface trick, a way to make the user feel a sense of abundance about the rewrapping of songs they already liked, taking algorithmically premeditated low-risk content and making the options seem new–it was, as ever, an illusion of choice.

I like the book, but it suffers from the same affliction a lot of academic-adjacent mainstream books fall victim to: it just sort of states a position with single line of attribution, often nested (as in, a citation to a work that cites the work that is being upheld):

Cornils points to Jacques Attali’s idea in Noise: The political Economy of Music that “the music industry is not devoted to the production of supply, but rather to the production of demand itself.” Playlists don’t just respond to the users’ musical interests, but manufacture them, too.

The endnote for this section is just an endnote. I’m not expecting Mood Machine to be 500 pages, but I do want a bit more. Does that make me greedy? Yes. This topic is really fascinating, and perhaps the book itself is positioned almost exclusively for a market segment that hasn’t read Spotify Teardown. In admittedly unfair shorthand, I sometimes categorize most non-fiction books on a scale ranging between academic or journalistic. Academics tend to write for depth, while journalists write for clarity. Mood Machine is good at the how and when and what, but when it gets to the why–which is honestly what I’m most interested in–the book lets you draw your own conclusions. Fair, again, for reporting, but sometimes we simply don’t get enough information on some of the more technical steps. Tautology and question begging build the rhetorical structures for assertions like “[P]laylists manufacture users’ musical interests” or “Data is never neutral–it reflects the biases of the culture and business logics from which it is generated” or “Algorithmic discovery was being used to bolster the corporate enclosure of music.” I want to append “…by doing [blank]” to each of these sentences, and a few paragraphs at least to each of these concepts.

Mood Machine is imprecise, vague–vibey, one might say–which makes it hard to know whether the it really is for the uninitiated, which the mile-high overview implies, or for the in-the-know, which the swift gloss might suggest:

The ambiguity of the term [chill] is perhaps by design, similar to how the industry use of the phrase “functional music” obscures how it usually just means soporific, playlist-friendly music, which is to say smoothed-out background music.

For whom it seems the book is writing is the mass of record-store regulars that have already turned against streaming, which is a fine market-segment to target but doesn’t exactly promulgate a wider understanding of what is actually happening under the sheen of the Spotify vibe factory. Like, what Spotify is doing can seem clever from the technical point of view, or even as a marketing move: applying stuff genre labels–content containers, really–onto specific sounds seems valuable, until you realize that artists most likely need to start playing toward the machine expectations. Horrifying. The semantic drift of “indie” from a tag that used to refer to an unsigned, independent artist into a profile for a listener who likes a certain soundscape is interesting, even if it is noxious to a community of artists that are no longer able to be called independent without assuming the role of feeding a consumer bespoke, decontextualized content:

“I don’t think a massive sound shift from an artist is well served by Spotify. Once you’re in the algorithm, it’s hard to escape where you fit into…The related artists are set,” [one independent musician] said. Content is more valuable when it is manageable, when it more legibly fits advertiser-friendly categorization. It discourages adventurousness.

And while I don’t need the author to be content-neutral, and certainly prefer the more humanist and collective perspective in reality that I see expressed on the page, it often feels like Mood Machine is afraid its readers might conflate being interested in what Spotify is doing for being in support of it. I am interested in the details of Spotify’s algo and marketing and cultural weight, even—maybe especially—because I think it stinks. Give me more. I know the author, as a music reporter, agree that it stinks, but I would like the author, as an author, to give me more detail.

Speaking of detail, the industry has to love the metrics and data that rigid labeling produces, but someone has to have realized that as the data starts to cycle, it’s just going to self-reinforce itself and smush out anything worth learning (see: Google search function; any and all large language model chatbots a half-decade from now). I, in fact, think it is pretty likely everyone has realized this data isn’t doing much but pointing back at itself, but since almost everyone’s job within the data analysis department of a company like Spotify depends on maintaining the fiction, it’s a lot like fiat currency–we pretend that the data scraped from listeners is valuable so that it retains its value. It is, actually, a lot more like Pizza Hut’s cost/value strategy: Spotify’s data also only has three ingredients: Cheese, Bread, and Sauce, and they simply need to find a new configurations–breadsticks, calzones, stuffed-crust–to sell you the same content in a different form. It’s really all just pizza, dude, but now it’s in a slightly different shape.

Unfortunately for me, the detailed exposé was the book I wanted. Credit where it is due: the subtitle was honest, even if I preferred reading the “Rise of Spotify” segment to the “Costs of the Perfect Playlist” portion. I did learn interesting things about modern Spotify that I had never heard or considered before: lower royalty rates are paid out in exchange for priority placement in ‘discovery mode’; unnamed artists create filler music on the cheap to puff up playlists and lower Spotify payouts; labels are paid out differently whether a song is streamed from a free or paid account. Mood Machine works through these details tentatively but never explicitly. It rubs against Spotify, but mostly at an oblique angle, from the vantage of musicians or employees interviewed. It’s a decent primer if you’re new to extractive tech or a burgeoning musician yourself, but if you already know your way around the landscape, there are more detailed maps to find bigger treasure available.