What happens when you outsource your sensibilities to the taste of the crowd? Two decades ago, a couple of artists attempted to do just that. In a fall 1995 show at the Dia Art Foundation in New York, the Russian-born American conceptualists Vitaly Komar and Alexander Melamid presented “The Most Wanted Paintings,” a project for which they used survey data to determine what elements Americans—and later Italians, Russians, Kenyans and many others—wanted to see in art, then composed “most” and “least” wanted paintings according to those criteria.
A few years later, Komar and Melamid teamed up with the composer Dave Soldier to apply the same principle to music, using input from 500 visitors to Dia’s early website. “The Most Wanted Song” was a smooth R&B duet that the composer statistically determined would be “unavoidably and uncontrollably ‘liked’ by 72 plus or minus 12 percent of listeners.”
“The Most Unwanted Song,” on the other hand, is more than 25 minutes long and features both bagpipe and accordion, which apparently people hate—they tie at 13 percent desirability—among many other instruments. It features elements of atonal music, advertising jingles and elevator music. In part, it’s an operatic soprano rapping about politics—and not in a Hamilton way. According to the composer, “fewer than 200 individuals of the world’s total population would enjoy this piece.” I find it fascinating, though I admit I’ve never made it all the way through.
Komar and Melamid’s songs project was an interesting early glimpse into how data might be used to mitigate listening experiences. But we’ve come a long way since then.
These days, my entire life is shaped by sophisticated machines that have a lot of information about my tastes and about other people’s. Amazon recommended my shampoo; Netflix and Hulu help me decide what to watch; OKCupid told me to date my boyfriend; he wooed me at a wine bar Yelp suggested he might like.
The less time we spend searching for something we’ll enjoy, the theory goes, the more time we have to enjoy it. And it’s a happy simultaneous development with the last two decades’ dramatic opening outward, like a time-lapse rose unfurling, of access both to tangible, buyable items and to creative work. We’ve never had easier access to every product, song and film we can think of—or better tools to narrow those ever-widening options down to the ones we’ll like best.
But how often do we stop to ask two ourselves two pretty basic questions about these recommendation genies: One, how do they work? And two, how are they affecting us?
Because these algorithms seep into every aspect of our waking lives, there’s no shortage of services we could examine to come to a better understanding of how they work. Spotify is a useful case study, both because of its popularity—it has over 140 million active users—and because its recommendation algorithm is particularly sophisticated.
Spotify combines three different technologies to create its insanely popular Discover Weekly playlists, which offer each of its users 30 new songs to check out every Monday. Better understanding these three separate mechanisms is key to understanding how Spotify—through its algorithm—is able to make you feel like it knows you better than you know yourself.
First up, collaborative filtering combines data about what you like and data about what people like you like. It’s sometimes known as the “Netflix model,” because Netflix basically built a business on it. Next is audio modeling, which analyzes the attributes of a song itself—aspects like timbre, pitch and tempo—quantitatively, using “deep learning” or neural networks (a bit like facial recognition).
And last are natural language processing models, which work by analyzing text, in a neat way: Scraping the web for what people are saying about music and the words they use to describe it. Using huge data sets created with that information, it creates associations between songs.
Spotify’s Discover Weekly—and lots of other algorithms like those used by Amazon, OKCupid, Netflix and Google Maps—have gotten very good at doing exactly what they’re programmed to do: recommend experiences to us that we will like. Or, at least, that we almost certainly won’t dislike.
From the perspective of the programmers who developed them, that’s a complete success. You like the songs, you buy the shampoo, you date the guy. You spend more time and more money on those sites. Surprising and delighting you, if they do—and they do, most of the time—were always just a means to that end.
But what about from the perspective of a human being who wants to squeeze as many experiences, as much adventure, as possible into the miserably short time we get to be alive? Is anything lost when what you encounter is all but guaranteed to be appealing?
For those of us who love music—or who love anything around which humans build communities—the thrill of discovery is related to, but distinct from, the pleasure of enjoyment. The risk that comes with optimizing the latter is eliminating the former. Think of how you first learned about the songs that made you love music: Was it an older sibling, pilgrimages to a record store, maybe a magazine subscription or MTV? How much chaff did you have to sort through before you found what moved you?
And what about the thumbs-downs: Is there anything useful about being exposed to music you don’t like? Has that ever made you learn something about music, or about yourself? Think of how a music service might sum up your “taste profile,” then think of the songs that don’t fit into it. Where did they come from?
I think it’s a dangerous practice to tidy up our experiences of something as complex and intimate as listening to music in the binary mode to which algorithms can necessarily be boiled down: 1/0; accept/reject. Recommendation services take for granted that giving a song a thumbs-up is simple, basic, knowable—but I’m not sure a machine really knows what it means, or even that we do.
So marvel at the recommendations, sure. But as you do, leave room for the stuff for which only human experience can hollow out a space in your heart.