I'm a happy subscriber to the AppleMusic service. It has almost everything I ever care to hear, AAC is a really high quality format that I find indistinguishable from lossless on high end gear and, at $10 per month for the individual or $15 for the family plan, I think it's a bargain.
Years ago, I ripped the thousands of CDs in my collection. When iTunes Match came online, I uploaded the collection there and think I may have played a physical CD about once since then. When AppleMusic came along, my library continued to grow with the a lot of new music I either found or that was recommended by Apple.
I enjoy many different musical genres and along with a large collection of classical recordings, my library has everything from hip-hop to bluegrass. My favorite genres and the ones to which I listen most often are blues, jazz, blues-rock and rock and roll. I listen to music daily and the genre decision basically comes down to what I feel like hearing at any given moment.
Sometimes, I don't feel much like thinking about what I'd like to hear and in the Music app on my iPad go to the "Fore You" tab to see what AppleMusic recommends for me. I enjoy listening to the AI generated Favorites Mix as I find it does a good job of picking songs that flow well together and it's this automatically generated playlist that got me thinking about the titular subject of this article.
I should add that this piece is being written by a guy who does not understand how AI or neural networks function in any but the most basic of manners. I should also add that I'm 100% confident that my findings did not come from anything Apple did intentionally but, rather, they raise questions about how an AI learns and makes decisions about one's preferences in this case.
My Music Library and Listening Habits
As I prepared for this article, I took a look at my AppleMusic library with classical and opera filtered out. Of the 3000 or so remaining albums, roughly 60% were recorded by African American artists. If one has a lot of jazz, early rock and roll and blues in their collection, they're going to have a lot of music performed by black musicians. I have something like 60 different Miles Davis albums, about 85 by John Coltrane, most of the Sonny Terry/Brownie McGee catalogue, all of Little Walter, loads of James Cotton and a ton more by African American artists who played the music I love.
I looked over my listening history over the past few months as I'd assume the AI that builds the playlist may take recently heard songs into account. My survey showed that I listened to roughly half black artists and half white over that period. I did go for a while while feeling very depressed listening to Lou Reed's Transformer and David Bowies Low over and over but the repetition is not reflected with Reed or Bowie being over represented in the playlist.
The Favorites Mix
About six weeks ago, I listened to what was then the most recent Favorites Mix. While listening from the top down, I noticed something: all of the songs at the beginning of the playlist were by white artists, a Jimi Hendrix song played and then most of the songs in the latter half of the playlist were by black artists with a few white blues players tossed in. I thought it to be a curiosity and didn't think much about it. I did the same the following week and observed the same: white artists first, a Jimi song, black artists to follow. A third week and I observed the same. For the past two weeks, I've taken notes and have found the following:
The Favorites mix that came out on April 24 started with Elvis Costello, Rush, Led Zeppelin, Rolling Stones, Dylan, Springsteen, Cream, Beatles, Patti Smith, Bowie, Lou reed, Yardbirds, Sparks and Zappa before the Hendrix song comes on. That’s 14 consecutive songs by white artists from a collection that’s mostly recorded by black artists. After these, there’s another Hendrix song from the blues and not rock and roll genre and then it’s all blues songs by both black and some white performers.
On May 2, it started with Dylan, Cream, Johnny Winter, Springsteen, Rolling Stones, Sparks, Rush, Led Zeppelin, Lou Reed, Jimi Hendrix, Zappa, Beatles, Elvis Costello and Patti Smith. Of the first 14, 13 are white artists and Jimi is near the point in which the playlist switches primarily to black artists.
When I'm listening to a Favorites Mix from the top down, I know that when I reach Jimi, I'm about to enter the black neighborhood in the land of the playlist. A couple of white blues artists will show up in the black part of the list but, excepting Jimi, I've haven't seen a black artist at or near the top of the list. It feels as though the black artists are being pushed to the back of the musical bus.
No Malicious Intent
Relatively recently, Microsoft released a Twitter bot based in an AI to explore computational conversation. Within 24 hours, the bot was spouting hateful, racist and sexist statements. Twitter taught the Microsoft bot to turn ugly very fast and MS had to kill the account.
In the music selection case, I doubt anyone was trying to make it be racist. As I said at the top, I don't know much about neural networks or how they learn. In a discussion with someone who does understand this stuff, I learned that my initial ideas on how this happens was hopelessly naive. I had thought I may have found a few factors that may lead to my favorite white artists coming before my favorite black artists. Apparently, this kind of thing is really complicated and unintentional biases in the preferences of millions of AppleMusic users may indeed be the root cause.
Ai and networks are still in their infancy. Learning algorithms are imperfect at best and, if my expert friend is correct, the number of factors that go into making a playlist like the Apple Favorite Mix is pretty big and I'm unlikely to figure it out on my own.
Of course, all of this could be a weird coincidence but five weeks in a row is pretty compelling that the AI wants to segregate my playlist into a white and black neighborhood with Jimi serving as the proverbial railroad tracks.