The Uncanny Venue
How Spotify is Making Us All Robot-Lovers
I said, “let there be bass.” And there was bass. I said, “here come the drums, I bet,” and sure enough, there were the drums. I said, “Holy shit, if this is another song with those goddamn vocals I’m going to lose it.” And lo, for the fourth or fifth (sixth? ninth?) song in a row, Hushed Indie Songsmith vocals dribbled out of my headphones like a spilled cup of overpriced coffee.
It was a couple of months after Spotify, the music-streaming company that has conquered the planet, debuted its Radio service, which takes a song or playlist and spits out a functionally infinite series of related, recommended tracks. I’d given it my 3,000-song, folk-and-psychedelia-heavy iTunes music library, to play with, and oh man, what it gave me back was bad: a totally monochrome feed of indie sludge, each boring song not just similar but nearly identical. Ten or fifteen minutes in it felt like you could see the precise variables the recommendation algorithm had set, because there were only a few — this tempo; drums that come in on this bar; quiet, doubled male vocals that sound like this; let there be bass. I turned it off.
Automated music recommendation is old news by now; Pandora has been serving up suggestions to anybody with an Internet connection since 2004. But now, Spotify is taking a new approach that deserves another listen. In July of 2015, it debuted a service whose impact has not only eclipsed Pandora (and, thankfully, its own Radio), but that is on track to change how — and why — people find new music online.
Every Monday, Discover Weekly, a customized, computer-generated playlist of about thirty songs, quietly slips into the accounts of the company’s more than seventy-five million users. From the outset, two things set these recs apart from Pandora and the rest of the pack.
The first is the tech. For its part, Pandora’s recommendations may seem smooth and automatic, but they’re powered by good old-fashioned analog human toil — the Music Genome Project, on which the service bases its suggestions, is a “taxonomy of musical information” based on “ten years of musical analysis by [a] trained team of musicologists.” In short, every song on Pandora is individually appraised by a person sitting in a room somewhere, screened for characteristics like tempo, genre, and dominant instruments, and plugged into a giant, handmade network of musical relationships before being delivered to your speakers.
That’s a hell of a lot of effort, and Spotify found a labor-saving workaround: Use the users. There are more than two billion listener-made playlists on the service, which means that odds are you have a musical twin somewhere on the site. Based on the “taste profile” the company has built of you out of your song searches, listens, and skips, Discover Weekly finds tunes that people like you have listened to but that you personally haven’t, then compiles those songs into a playlist. It’s Amazon’s “customers like you have also bought…” feature applied to music discovery.
More interesting than the tech, though, is the philosophy that guides it. Despite all the human effort that goes into it, music recommendation on Pandora (or Spotify’s own misbegotten radio feature, for that matter) is at its core an artificial experience. It’s quintessential Web 2.0: You go to a dedicated site or app and plug in what sorts of things you want to hear, and a far-off server clicks and whirs and feeds you back what you asked for. You can even share your finds on Pandora’s social feed.
Discover Weekly’s a different paradigm entirely. You don’t need to go anywhere for it — it just arrives in your sidebar, ready to listen when it’s convenient, like a mixtape shipped to you by an old friend. And it does so in the form of a playlist, a digital mix-tape that could have been made, in terms of both available tools and content, by any person on the site. It makes automated recommendations, in a word, more human; nobody ever lovingly crafted their high school sweetheart a Pandora station.
Making automated recommendations feel more human isn’t a cutesie metaphor — it’s the explicit goal of the product, one that’s soaking into the company zeitgeist. A Fast Company interview with Discover Weekly’s man-in-charge, Matthew Ogle, described his creation as “a playlist of songs designed to feel like it was handcrafted for you by a good friend.” The Verge’s Ben Popper described the experience of taking Discover Weekly recommendations as receiving “an intimate gift from someone who knew my tastes inside and out, and wasn’t afraid to throw me a curveball.” Recommended music radio helps you keep a look out for new music; Discover Weekly is there to look out for you.
That change in language and presentation is way more than the PR spin it sounds like; Spotify is barreling with great intention into a potentially dangerous place for the company to be: a space where the line between humans and computers is blurred.
People have been playing with blurring the lines between computers and people since Alan Turing, the father of modern computing, named his Enigma-decoding proto-computer Christopher, and as machines have started to clear the technical hurdles to acting believably human they’ve run into an emotional one: the Uncanny Valley.
The Uncanny Valley, a phrase first coined by robotics professor Mashahiro Mori in 1970, is the feeling of discomfort you get when you interact with something that appears a little bit too human, but isn’t — think of the animatronic performers at Chuck E. Cheese’s, the skin-crawling weirdness of the CGI in The Polar Express, or the dead-eyed stares of the bots in the first Mass Effect game.
It’s an innate, physical feeling with a discernable cause: A 2013 study found that the anxiety and creepiness of the uncanny valley comes from the interactions you experience not lining up with your brain’s expectations. Watch a robot jerk and jolt its way toward you toward you and you’ll feel nothing, because that sort of movement is what you’ve been conditioned to expect from a robot; but make eye-contact with a smiling, jerking Repilee Q2 and it’s a different beast entirely.
Strictly speaking, the Uncanny Valley refers to things you principally see or socially interact with — a character in a computer game or an android — but as people have put more and more of their lives online its definition has expanded to mean when a computer knows or does something it just shouldn’t.
Case in point: Facebook declared February 4th “Friend Day” and automatically generated a cute video of each user’s photos with their friends. It was meant to be charming, but in some circles of the internet it was on-the-nose enough to just be weird. “I find it deeply creepy that a Facebook algorithm can so easily provoke genuine nostalgia and emotion in me,” said MSNBC host Chris Hayes in a pair of tweets reacting to Friend Day. “Like, the emotional version of the Uncanny Valley.”
All of this is to say that, by aiming to make its automated music recommendations believably human, Spotify is entering well-traveled and treacherous waters. It’s entering a battle with its own, auditory version of the Uncanny Valley; call it the Uncanny Venue.
The danger for Spotify comes in part because its foray into person-like recommendations is working, as those Fast Company or The Verge profiles can attest. But making an auto-generated playlist feel like a good friend can make the sense of betrayal that much stronger when that good friend screws up.
“Having an algorithm decide that I would like this song is the most hurtful thing a computer has ever done to me,” wrote Fran Hoepfner in “My Spotify Discover Weekly Playlist Thinks I Am A Garbage Person,” an article about Discover Weekly recommendations gone wrong. It’s melodramatic, sure, but the sentiment is built on two pretty strange ideas: that a computer-generated song suggestion should carry emotional weight, and that that feeling is held widely enough to write an article about. (“Five Times I Had to Press Thumbs-Down on a Pandora Station” didn’t get quite as many shares.)
But despite a human-emulating approach that has given people the creeps in other fields and the risk of alienating users, Discover Weekly is exploding: its playlists generated 1.7 billion listens in the five months after its release.
Spotify took explicit steps to make its technology appear as human as possible and got its users to emotionally invest instead of running away. In other words, it crossed the Uncanny Venue.
Our comfort with Discover Weekly says a lot more about us than it does about Spotify. Could it be that music recommendations are too abstract a thing to set off Uncanny Valley alarm bells? Or that Discover Weekly is still sitting pretty on the robot side of things? Regardless, the fact alone that a company as valuable as Spotify would even attempt to couch a totally digital product in human terms underscores our increasing comfort with dropping the veil between human and machine.
After all, most Amazon shoppers have reconciled themselves with the fact that a server can know more about their shopping preferences than their closest family. Or take XiaoIce, a chat-bot popular in Asia that has gotten really good at talking with people — almost human-level good, but not quite. But instead of disconcerting users, Uncanny Valley-style, or quickly getting repetitive and boring a la Cleverbot, XiaoIce has thrived — it boasts forty million users and chats that go back and forth twenty-three times on average; there are many text conversations between actual humans that can’t say as much.
Ultimately, what is or is not in the Uncanny Valley comes down to expectations, and those expectations are changing for humans and machines alike. Just as people are allowing computers larger and larger roles in their daily, physical lives, so too are they letting them into their internal, emotional lives. That’s why an almost-human chat app can take off: If a computer can schedule my meetings and set my home temperature, why can’t it shoot the shit?
But Discover Weekly is special because it’s not about humans giving computers emotions — it’s about computers giving humans emotions. Music is the quickest, easiest way to manipulate people into feeling something on-demand; ask any movie score composer or baseball fan during the national anthem. We’ve dangled the keys to our emotional state. Discover Weekly has snatched them and used them to serve up a weekly dose of feelings that its users, for the first time, have completely ceded to a machine, and 1.7 billion listens’ worth of people are fine with that. It’s another stitch in the increasingly complicated and fuzzy relationship between people and machines, the physical and digital space, and it’s harder than ever to tell who’s really at the controls. That’s not implicitly good or bad, but in a way the moral implications are hardly relevant, because they changes are happening and they’re here to stay.
And by the way, the one song that convinced me of Discover Weekly’s potential and turned me into a die-hard listener? It was Marina and the Diamonds’ I Am Not a Robot.