Artificial intelligence as an artist?
This is not the first experiment where creators have tried to use algorithms to create something like a work of art. However, while the recently described Czech experiment with theater art, created using the GPT-2 text generator, looked very credible and was well documented, the initiative described here has its origins rather from a meeting of an advertising agency.
On the project’s website, the creators did not give us a chance to find out what type of algorithms were used, or how much it actually contributed to the creation of the whole. After digging through several music and technology websites, I finally found an article in which the creators confirmed that a program from the Google AI stable called Magenta was used for this.
The “feeding” process was such that the elements of individual songs were broken down into simple sequences stored in MIDI files. The same was done with the texts that were given to the algorithms in a pure form, not related to the music.
Secondary? OK but flawless?
After listening to the songs on the website that imitate the works of Jimi Hendrix, Jim Morrison and Amy Winehouse, two things catch your ears. The first is the repetition of the songs placed there, which sound like a typical work of a craftsman modeled on the songs of famous artists. This is something to be expected from both human and algorithms. The second thing is the lack of imperfections characteristic of algorithms, such as errors, repetitions or surprising musical combinations of “blocks”, which proves that the contribution of AI was negligible here.
This type of primitiveness of today’s solutions in the field of AI is perfectly shown in the aforementioned show “AI: When a Robot writes a play”. There, the creators made every effort to reflect the nature of the algorithms’ work as accurately as possible, and the methodology of interference was openly described on the website and told at a press conference. In general, efforts were made to force the AI to generate successive iterations of the text, so that the final work was as much as possible the work of a “machine”. The creators of the experiment with essays written by GPT-3 did a similar thing.
In the case of this album it looks like that after feeding the algorithm with the mentioned MIDI sequences and texts, the returned fragments were probably nothing more than distorted variations of the batch and of questionable quality. As the creators revealed to The Rolling Stones magazine, 90% were fit for the trash. From the rest, the proper artist, undoubtedly belonging to the genre of homo sapiens, chose what he could and adapted to his own composition.
This means that the contribution of algorithms to the creation of these works is virtually none. You might as well let the children ruffle on the synthesizer and pick out a mass of nonsensical sequences that will fit into the piece planned by the human composer.
The science of counter marketing
I have the impression that more and more technological topics are absurdly used by marketers to promote their own initiatives. Perhaps there would be nothing wrong with that, if the image at what level in this respect we are today was not distorted in such a way. Today’s AI is not so much unable to write a flawless song, as it cannot even write a terrible song. Such projects, in which the “nature” of the algorithms used are completely lost, are, in a sense, a fraud from the point of view of a person interested in this subject.
Sources: , , 
More from the category Artificial intelligence: