Can a computer program ever compose music that’s as good as what a first-rate human composer can devise? I claim that the answer is “no.” And I’m going to prove it.
I started thinking about this question after reading a new piece about David Cope, the retired UC Santa Cruz professor who has devoted his life to developing software that composes music. Cope has gotten a fair amount of press over the years — quite a lot, really, considering the marginal quality of the music his programs have produced. He has also garnered a fair amount of criticism. Since I’m one of the critics, I felt not only curiosity about what he’s up to now, but a sense of obligation. I don’t want to be unfair to the guy, after all. Maybe my view 20 years ago was too harsh.
The new article (you can read it online at http://www.miller-mccune.com/culture-society/triumph-of-the-cyborg-composer-8507/) was written by Ryan Blitstein. Blitstein’s bio identifies him as a journalist, but does not mention any musical training or musical credentials, a fact that might alert us to potential weaknesses in his presentation. In what follows, I’ll weave back and forth between Cope’s own work as I understand it and the claims made for it by Blitstein. Cope’s own web pages are at http://artsites.ucsc.edu/faculty/cope/index.html.
In a nutshell, Cope has written several programs over the years that analyze musical materials fed into them (melody lines, chord progressions, rhythms, polyphonic textures, voice leading, and so on) and then produce new music that conforms to or arises out of the analysis. In his early work, for instance, he input hundreds of four-part chorales by Bach. After processing this data, the software was able to synthesize entirely new four-part chorales that were more or less in the style of Bach.
Not having studied exactly how the program goes about its analysis, I’m not qualified to judge how good it is at doing that. Nonetheless, as an academic exercise that may prove useful to musicologists, it’s a remarkable achievement of software design. This may be Cope’s real contribution to the universe of discourse.
The claim is repeatedly made, perhaps by Cope or perhaps by those around him (I don’t have any direct quotes from him on this subject) that listeners who are not aware that the music was composed by a computer can’t tell. They believe, or assume, that it was composed by a human being. Or so we’re told.
It may be true. But we may want to remind ourselves that until recently most people believed the Earth was flat. Human judgment is notoriously prone to error. In any case, I’m not aware of any double-blind studies in which listeners were asked to pick which composition was written by a human and which by a computer. That would be the only valid way to determine once and for all how good the program’s output is. And of course both compositions would have to be played, without inflection, by a synthesizer, because a human performer might easily introduce various sorts of bias.
But would such a double-blind study be valid? The picture is actually quite a bit murkier than even such a supposedly scientific test would be likely to illumine.
To begin with, a computer cannot analyze musical compositions on its own. The computer has to be taught to do this. Before it can discern a single feature of a composition or body of compositions, it has to be taught, in some manner, how to recognize such a feature. Features that it is not taught to recognize will remain invisible to it. The computer’s teacher is, in this instance, David Cope. Musical features that exist, but that Cope doesn’t recognize as significant, won’t be encoded into the algorithm.
For instance, if Cope were to fail to recognize that the movement of a voice by leap rather than by step was a significant component of style, then the computer would choose leaps and steps purely at random. I’m not saying his analysis is that crude; I’m sure he did think to include an analysis of stepwise motion! But subtler issues may nevertheless have eluded him.
Thus, in the end, the software’s compositional ability rests entirely on David Cope’s musical understanding. In a double-blind study, we would not be listening to a computer’s compositions: We would be listening to David Cope’s instructions on musical analysis, which were being utilized in some manner by the computer. In addition, Cope would be choosing (most likely according to some unconscious criteria) which pieces to feed into the computer’s database for analysis and which to omit. Again, the human element can’t really be abstracted from the process.
Beyond this, for a double-blind study to be valid, we would have to choose at random from among the computer’s output scores. If we let David Cope or anyone else choose the very best or most successful outputs of the program for comparison with Bach, Mozart, or whoever, the results of the test will be wildly skewed.
This hand-picking of the outputs is, I believe, something that Cope does. I don’t have the impression that he’s releasing randomly chosen pieces from among the computer’s output. I could be wrong, but I would need to see the evidence.
Cope’s software, Blitstein writes, produced “scores so impressive that classical music scholars failed to identify them as computer-created.” The scholars are not, however, identified by name. Nor are the conditions under which they were bamboozled described. It does make a difference.
Next, Blitstein’s article gives us this: Cope’s “creation,” by which Blitstein presumably means his software, “raised troubling questions: If a machine could write a Mozart sonata every bit as good as the originals, then what was so special about Mozart? And was there really any soul behind the great works, or were Beethoven and his ilk just clever mathematical manipulators of notes? Cope’s answers — not much, and yes — made some people very angry.”
The difficulties with this passage are manifold.
First, there’s an enormous gap between failing to identify a composition as computer-created and claiming that it’s “as good as Mozart.” The sonatas of Dussek, Hummel, and Cramer were indisputably created by humans, but not many listeners would claim they’re as good as Mozart. For all we know, one of those scholars said, “Gee, I can’t tell if that’s a computer composition or if it’s something scrawled out by a third-rate student of Dussek’s who later found happiness as a coal miner.”
Second, “what was so special about Mozart” was that he was the person who invented or refined the musical style we identify today as Mozartean. Until a software program invents a new and original style that is as compelling as Mozart’s style, the score will still be humans 1, computer algorithms 0.
Third, “ilk”? Shall we load the argument with a gratuitous insult? Oh, yes, let’s.
Leaving aside the question of what a “soul” might be (the term has, as far as I can determine, no meaning whatever, a point on which David Cope and I agree), it’s certainly true that the great classical composers were “clever mathematical manipulators of notes.” Nobody disputes it. The difference between a human manipulator of notes and a silicon manipulator of notes is this: The human has, at the beginning of and throughout the compositional process, an end in view — a goal. The goal may be to please audiences and thereby make money (a constant concern for both Beethoven and Mozart, by the way). It may be to express some deeply felt emotion. It may be to impress a possible romantic partner with one’s technical virtuosity. Or the goal may be something entirely different.
Perhaps these ulterior motives are what Blitstein means by “soul.” In any event, a computer algorithm has none of these motives. It is never more than a clever mathematical manipulator of notes. And that’s the difference. If David Cope actually asserted that Beethoven and his ilk were “just clever mathematical manipulators of notes,” then David Cope is seriously out of touch with reality. But we don’t know that he said that. We only know that Blitstein alleges that he said it.
Next we learn that Cope’s 1990s-era software system has been eviscerated. The program still exists, but Cope himself got rid of the database of musical analysis. “For a time,” Blitstein tells us, “such condemnation fueled his creativity, but eventually, after years of hemming and hawing, Cope dragged [his analysis database] into the trash folder.” What’s interesting about this paragraph is what it doesn’t say. It’s clearly written in such a way as to imply that it was Cope’s discouragement over the criticism that his work received that caused him to delete the data. But Blitstein doesn’t actually say that the “hemming and hawing” — which is in any case not the correct phrase; the correct term would be “shilly-shallying” — was caused by the criticism. It may have been caused by Cope’s own dissatisfaction with how the software was working. Or by clinical depression, for all we know.
Having finished with his survey of the recent past, Blitstein turns to the real subject of the article — Cope’s new software, dubbed “Emily Howell.” This program, Blitstein tells us, “aims to do what many said [the earlier software] couldn’t: create original, modern music.”
We may want to try to understand what Blitstein means by “modern music.” The two mp3 pieces he provides by way of demonstration, both of them performed on solo piano, are thoroughly tonal, and thus not part of the mainstream of modern classical music, nor anywhere close to it. The first piece ripples along rapidly, rather in the manner of a Chopin etude. It has a few surprising and very pleasant chord progressions, but no melody ever appears, and the texture never varies. It sounds, frankly, computer-generated. It has no meaning, and thus no soul. The second piece is plainly based on the model of a Bach fugue — one of the slow ones. The fugue subject is very, very long and meandering, and when the second voice finally enters, it’s not on the dominant (which Bach would have done as a matter of course) — it’s on the tonic, just like the first voice. So even as a realization of a Bach fugue, the piece is a failure.
Blitstein then states, baldly, that “Emily Howell’s” compositions “are innovative, unique and — according to some in the small community of listeners who’ve heard them performed live — superb.” I would like to have read quotes from some of these listeners myself. I’d like to know who the listeners were, and what their musical backgrounds are. Bad imitation Bach and bad imitation Chopin would hardly seem to qualify as “innovative” or “unique.” But again, we can’t lay the blame for these claims on Cope. Blitstein is the one who is doing the gushing.
Blitstein tells us straight out that Cope is “exposing revered composers as unknowing plagiarists.” That’s an interesting charge, though it’s not supported by a single example of a revered composer who plagiarized, knowingly or otherwise, nor by a quote from Cope giving us an example of plagiarism. But I can offer one. Beethoven’s Ninth Symphony includes a Turkish march, which was a popular genre at the time. That’s certainly a case of plagiarism. It doesn’t diminish the value of the Ninth Symphony, however.
I believe it was Stravinsky who said, “Bad composers plagiarize. Good composers steal.” By which I think he meant that all composers borrow compositional devices from one another, but that good composers use the borrowed devices in fresh and musically insightful ways, as Beethoven unquestionably did, while bad composers just rehash what they’ve borrowed without really understanding the underlying musical impulse or adding anything new.
Borrowing from previous composers without understanding the underlying musical impulse and without adding anything new is, of course, precisely what Cope’s software does. That’s what it’s designed to do, and that’s all it can ever be designed to do.
Blitstein then turns to a discussion of how Cope initially developed his software (not “Emily Howell” but the earlier program that he has since deprecated). At this point, the reductionist flimflam that I believe Cope himself has espoused, at least implicitly, takes center stage. We’re told that Cope developed an entire “grammar of musical storytelling,” with which to analyze existing scores algorithmically. “Finally,” Blitstein tells us, “Cope’s program could divine what made Bach sound like Bach and create music in that style…. It was as if the software had somehow captured Bach’s spirit — and it performed just as well in producing new Mozart compositions and Shakespeare sonnets.” Whoa! Shakespeare sonnets! That’s the first time I’ve heard that claim made for Cope’s software.
What’s especially interesting about this claim is that, if it were true (and of course anyone can see at a glance that it couldn’t possibly be true, without even having read any of the supposed sonnets), then Shakespeare would have been just a clever mathematical manipulator of words. Does anybody but Ryan Blitstein believe that? If you do, raise your hand. You, Stevie? Oh, you just want to go to the bathroom. Go ahead, then.
Finally, we get to something resembling an actual interview with Cope. “In his view,” Blitstein says, “all music — and, really, any creative pursuit — is largely based on previously created works. Call it standing on the shoulders of giants; call it plagiarism. Everything we create is just a product of recombination.” Well, yes, that’s undeniable, though Blitstein blithely slips a cog between “largely based,” which is true enough, and “everything,” which is bullshit.
You see, things change. Novelty enters the system at many, many points. Innovation is a constant in art. When Mozart sat down at the keyboard, all of the chord voicings used by Bill Evans and McCoy Tyner were right there in front of him. But he couldn’t see them. They were invisible, because they hadn’t been invented yet. At some point, each of those voicings was played for the first time by someone (Evans or Tyner, or someone who preceded them).
Yes, a computer can innovate, if by “innovate” we mean generating new ideas at random out of the dust of disassembled old ideas. But human artists don’t generate ideas at random. We do so for emotional reasons, in order to create emotional effects. You can see this in Tchaikovsky, for example. Blitstein tells us about Cope’s reaction, as a child, to Tchaikovsky’s “Romeo & Juliet” overture. One of the interesting things about that overture is the repeated falling motif (in the French horns, if memory serves) beneath the romantic melody. Tchaikovsky used this type of motif in other works as well (the “Francesca da Rimini” overture, for instance). To him, it evoked tragic feelings. That was why he used it. It’s a sigh. But a computer doesn’t know that it’s a sigh — unless a human operator tells it that.
As human listeners, our reaction to that type of motif is probably instinctive. A basic knowledge of common practice harmony is helpful, but beyond that, we don’t need to be told that Tchaikovsky is sighing. We listen, and we get it. A computer can’t get it, because a computer has no instinctual apparatus with which to get it.
Blitstein misses another crucial point as well. What’s interesting about creativity is not the raw fact of recombination, which is inevitable, but the manner in which it occurs. Recombination, in my experience, is largely unconscious — and it’s based not on a linear jumbling together of elements in random fashion but on a deep understanding of the grammar and syntax of music. Grammar and syntax are used to create musical statements that have meaning. The meaning may be abstract, or it may be more concrete, but it’s there.
“Cope believes,” Blitstein says, “that music comes from other works composers have heard, which they slice and dice subconsciously and piece together in novel ways. How else could a style like classical music last over three or four centuries?” You’ll note the reductionism of the term “slice and dice,” which entirely ignores the questions of musical grammar and meaning. It reduces the process to random jumbling. In addition, Blitstein seems to think that “classical music” is “a style.” He’s jamming Palestrina and Bartok into the same bottle — but he’s a journalist, not a musician. God knows why he got the assignment to interview David Cope. They could have asked me to do it. Hell, I might even like the guy. He’s certainly brighter and better trained than I am. We both like software, and we agree about the soul. But I digress.
Here’s a direct quote from Cope, discussing the compositional process used by humans: “Everybody copies from everybody. The skill is in how large a fragment you choose to copy and how elegantly you can put them together.” This is surely a cynical and simplistic view of what composing is, though perhaps it’s not an entirely surprising view, coming from a man who turned to computer-assisted composition due to a severe case of writer’s block. And what is that word “elegantly” doing floating around in there? Has “elegance” been encoded in the software algorithm, or hasn’t it? If it hasn’t, then the algorithm has failed. If it has, I’d sure like to know how.
I wasn’t aware that Cope had been written up by Douglas Hofstadter, but Cope’s description of a conversation with Hofstadter, as reported by Blitstein, is revealing enough to be worth quoting at length:
“As Cope sees it [Blitstein writes], Bach merely had an extraordinary ability to manipulate notes in a way that made people who heard his music have intense emotional reactions. He describes his sometimes flabbergasting conversations with Hofstadter: ‘I’d pull down a score and say, “Look at this. What’s on this page?” And he’d say, “That’s Beethoven, that’s music of great spirit and great soul.” And I’d say, “Wow, isn’t that incredible! To me, it’s a bunch of black dots and black lines on white paper! Where’s the soul in there?”‘
“Cope thinks the old cliché of beauty in the eye of the beholder explains the situation well: ‘The dots and lines on paper are merely triggers that set things off in our mind, do all the wonderful things that give us excitement and love of the music, and we falsely believe that somewhere in that music is the thing we’re feeling,’ he says. ‘I don’t know what the hell “soul” is. I don’t know that we have any of it.'”
This passage begins by setting up a false dichotomy: Either music is nothing but patterns of dots, or it’s an expression of “soul” and “spirit.” I happen to agree with Cope that “soul” is a meaningless term, but I’m not going to be sucked in by that dichotomy. There is a vast gray area between the two extremes. An important part of the gray area is intentionality. Beethoven had intentions. He intended that a piece of music be understood in a certain way. It’s entirely possible that Hummel or Dussek had similar intentions, but they failed, because their ability to synthesize large musical statements in ways that conveyed their intended meaning was inferior to Beethoven’s.
Cope seems quite clearly to be saying that if we perceive Beethoven’s music to be more profoundly meaningful than the music of Dussek or Hummel, it’s because of our reactions to the music, not because of anything inherent in the music itself. And that’s just plain wrong.
Tchaikovsky was sighing. He was sighing over the tragedy of the human condition, as the story of Romeo and Juliet reveals it to us. As a child, David Cope was profoundly affected by the “Romeo & Juliet” overture, but as an adult he has reduced what he felt to “a bunch of black dots and black lines on white paper.” And that’s profoundly sad.
When we listen to the music produced by Cope’s software, what we hear, very plainly, is a bunch of black dots and black lines on white paper. There is no ghost in the machine. And no matter how complete the analysis or how thorough the database, there will never be a ghost in the machine. If listeners are fooled, then maybe in a technical sense the composition program (a future descendant, perhaps, of David Cope’s work) will have passed the Turing Test. But that tells us only that there’s something fundamentally wrong with the Turing Test.
If a computer tells you the Earth is flat, will you believe it?