Thursday’s Google Doodle celebrated the 334th birthday of famed composer Johann Sebastian Bach, with a twist: It was the first Doodle to incorporate machine learning. Users could create a melody, then the Doodle would automatically generate custom harmonies to produce a full composition in Bach’s style. It was delightful for many Google users, but it also stepped into a controversy that has been brewing in musical circles for years.
Google says in its “Behind the Doodle” video that it chose Bach’s music as the subject of the first A.I. Doodle because he had a characteristic style and composed with a set of musical rules in mind. This formulaic quality made his work an ideal subject for a machine-learning algorithm to train on. But Google is far from the first group to undertake this type of musical A.I. hybrid project. It’s something that alarms some music academics and composers, who dislike that an algorithm can distill whatever music it analyzes into mechanical tropes, without being able to fully capture the creative, thoughtful qualities of human-composed art.
The results of the Google Doodle were also particularly grating to Bach specialists. The algorithm made a lot of mistakes that even a first-year music student learning about counterpoint (the set of rules Bach and others used in composition) would know to avoid. “I found the Doodle to be a bit of an odd tribute to Bach because the results bore so little resemblance to Bach’s style,” says Christopher Brody, a Bach scholar and music theorist at the University of Louisville. “It would be hard to think of a single generally true fact about Bach’s idiom that the A.I. seems to have accurately observed and manages not to break almost constantly.” Christopher White, a music theorist who studies algorithmic and computational approaches to music at the University of Massachusetts, Amherst, points out that a machine-learning algorithm like the one in the Doodle is generally composing off of hybrid rules it extrapolates in its own way, rather than the formalized concepts that humans study as the underpinnings of learning composition.
At Google, the person who led development of the Doodle’s machine learning algorithm is Anna Huang, an A.I. specialist who originally trained as a composer. To her, the Doodle attempts to mediate the inherent tension between creative musical development and the more formulaic attributes of machine learning. “As a composer, I find myself operating in different modes,” she told Slate. “Sometimes I want to work out an idea very precisely in all its details, other times I write down notes just as a placeholder so that I can see the overall shape of the piece. Sometimes I hear an idea from within, sometimes I wish there were raw material that I could sculpt. Machine learning can offer us ways to explore the spaces in between.”
Skeptical academic music theorists had fun experimenting with trying to stump the Doodle’s algorithms by feeding in bizarre melodies from experimental 20th-century music. They even tried inputting a melodic snippet from Usher’s “Yeah.” (That one was actually pretty cool.)
When I saw the Bach-lash begin, I decided that as a music theory researcher and teacher, I needed to plug in my own melody and grade the machine’s efforts.
The melody I fed to the Doodle was meant to be in the key of C major, but the Doodle produced a composition in A minor. This key is closely related to C major, so I understand how the Doodle may have misinterpreted it. But the harmonization was funky from the start because I hadn’t put any “A” notes in my melody, meaning the composition ended up feeling disjointed, as if two people had written the melody and the harmonies without fully communicating.
The Doodle also made a number of basic Bach faux pas, like adding in the voice leading error known as “parallel fifths and octaves.” This happens when two of the parts move in the same direction by the same interval. The Doodle also produced a composition that had overlaps between voices, another no-no in Bach’s baroque style. These errors might go unnoticed to a non-musician, but to those trained in Bach’s style, all the errors in the example I created sound like nails on a chalkboard.
But then I realized: So what? I teach Bach every year to students who struggle to emulate his style and precision. I had cringed at the idea that a computer could replace years of intensive composition study to generate great music. I noticed, though, how many people outside of musical communities—on Twitter and friends outside my academic circle—seemed to engage with the Doodle. The Doodle was designed so anyone could play around with it and plug in some notes, even if they aren’t familiar with music notation. And they would produce a lovely little composition that wasn’t always technically correct for Bach’s style, but sounded nice. This is a rare democratization of classical music that has value in its reach.
It’s still a shame, though, that there were so many errors packed into the short excerpts.
Since most people wouldn’t recognize them, it means that what might be one of their only experiences with this style of composition was not really representative of Bach, despite Google’s attempts to emulate the baroque master.
Academic music communities will continue to spar about the value of using machine-learning techniques to create music. But the reality is that these tools exist and appeal to lay people. If these technological developments can create opportunities to collaborate and expand academic discussions beyond conservatories and ivory towers, that’s something the community should celebrate and welcome.
Future Tense is a partnership of Slate, New America, and Arizona State University that examines emerging technologies, public policy, and society.
