Ludwig van Beethoven died in 1827 and left an impressive musical legacy. During his not very long life he composed 722 pieces of very varied length and character, which include more than 35 sonatas, 16 works for string quartets, numerous concerts for various instruments, masses, cantatas, an opera (Fidelio) …
But, without a doubt, his most famous works are the nine symphonies that he created between 1799 and 1824. Shortly after the genius died, his friend and personal secretary in his later years, the Austrian violinist Karl Holz, revealed that the artist born in Bonn in 1770 had been working for a while on what was to be his tenth symphony (commissioned by the Royal Philharmonic Society of London), and he had even played some of the first movements on the piano.
Unfortunately, only fragments of that unfinished work have remained in the composer’s notebooks , and a few loose references in his correspondence. In 1988, the British musicologist and composer Barry Cooper tried to reconstruct the symphony from these sketches, taking inspiration from the artist’s other compositions, but there is consensus among critics and experts: the great composition that the German musician was preparing would probably have sounded in a very different way from this worthy recreation.
On October 9, however, the world was able to hear an Alternative Tenth Symphony. Like Cooper’s, it has been created from the fragments and descriptions that Beethoven left behind, but the difference is that in this case, the work has been completed thanks to artificial intelligence (AI).
The American company Playform AI invented a machine learning system – a branch of artificial intelligence that develops techniques for computers to learn – which took as a reference the style of the 722 pieces composed by the German artist. On this basis, this AI composed a work that fits in with the natural progression of the rest of the Beethovian compositions. Written like this it seems almost as simple as pressing a button. In fact, to arrive at this new version has required the work of a great team of musicians, academics and computer engineers, led by Matthias Röder, director of the Karajan Institute located in Salzburg (Austria).
The biggest problem they faced was that the material that Beethoven left behind was, in addition to being scarce, disjointed. The best AI routines are capable of doing a decent job when it comes to recognizing patterns and continuing them. In music, this means that if you give an AI a nearly finished song, it can figure out what the final seconds will look like quite accurately. Creating something bigger out of nothing is a challenge of greater magnitude, and it cannot be considered just the result of applying artificial intelligence algorithms. Throughout the process, the human team has made decisions about which pieces could fit and which could not.
Unquestionably, the result sounds like Beethoven. To test it, the engineers responsible for the project devised a version in reverse of the Turing test, the famous test used to find out if software can give answers that are indistinguishable from those of a human being. In front of several musicians, music critics and academics versed in Beethoven’s work, they interpreted a few fragments of the tenth and challenged them to identify where the notes that he left written ended and where those that intelligence had extrapolated began. They couldn’t do it.
In the eighteen months that the work has lasted, the team has been able to orchestrate two movements of 20 minutes each. Like Cooper’s interpretation, it is likely that this tenth symphony was not exactly the one that sounded in the head of the creator of Fidelio, but we can use it as a starting point of a debate already known, but that is getting complicated: can it create works of art an artificial intelligence? The question has been posed on several occasions, but each time in less academic and more practical contexts. In 2018, for example, the Parisian art collective Obvious put up for sale through the Christie’s auction house several works created from machine learning routines.
Part of the difficulty in answering this question is that the definition of art is complex even for human beings themselves, who differ on exactly what it is. From the outset, we could rule out that an AI is capable of inventing artistic works, because in its creative process there is no emotional component, in theory one of the distinctive features of art.
But that does not mean that the productions of an artificial intelligence cannot have an artistic or imaginative touch. Scientists are still studying how creativity works in the human brain, but there are some known traits of creativity that are included in machine learning routines. We know, for example, that creativity depends to some extent on previous experiences. We always create under the influence of previous inventions and works, much like how a machine learning algorithm builds the models that it guides to create.
And another question arises: in the event that the work of an artificial intelligence can be considered art, shouldn’t the recognition fall on the people who programmed it? It makes sense, but there are those who argue that, then, the same consideration should be had – at least up to a point – with the teachers who have taught artists to paint.
What is clear is that artificial intelligence lacks two fundamental elements in this matter: will and free will . No program decides by itself that it wants to paint a picture or finish a symphony. It does it by order of a human and because it has been designed for it. It is the reason why the team that has worked on Beethoven’s Tenth Symphony considers the role of an AI in the art world to be that of a tool, such as a paintbrush or a violin, only much more advanced.
