AI creates copies of certain styles or songs without using their originals as training material … we call it MusicGST = Musical Generative Super Training.
Great attention around the last Beatles song supplemented with AI, and then Sir Paul says: “Don’t get excited, the new song contains nothing artificial.” Lucky again … but the discussion that AI generates songs in the style of stars on the basis of the original training data is getting bigger. One thing is clear: One thing is clear: the original hits and catalogs of the stars may only be used with the permission of the rights holders.
Fair enough, but to stay with the Beatles, for example: One doesn’t need the original albums and recordings of the Fab Four to create new music in their style.
Free or sufficiently otherwise available knowledge with which I feed the machine is sufficient for this:
- To “copy” the sound component, there are descriptions and sound examples of the respective recording period, of Höfner basses, Rickenbacker guitars, Echolette and Vox amps, Ludwig drums, etc., and from this you can compose great sound-(alt)-text-pairs automated for the AI.
- For each Beatles song and album there are countless reviews and texts, up to musicological analyses in the net, which give perfect music theory input as fodder for the machine.
- Of course, the Beatles would address the current issues of everyday life and love in their lyrics. So we can just take ChatGPT and augment the prompts with some info from the music theory corpus for song structure, complexity, or range, and wham: we’ll get lots of variations for possible lyrics in the desired style.
- For the vocals, we’ll build a small database of McCartney and Lennon voice doubles, because that’s what the case law says: people who look similar or have similar voices simply have to be “held out” by the original, even regardless of satire.
The consolidation of the four data components is automatable as a process for every style of an artist or an epoch and thus reproducible. There are enough secondary and primary information available for successful works no matter in which medium, whose combination with some training will create amazing results … copies. Of course it’s not allowed to use the name of the artist for this new creation as a selling point, because that would be fraud and there are enough legal remedies for that. But just to describe successful music periods, to create “soundalikes” or to process parts of the original works in new songs creatively in today’s context, for that the AI needs the catalog of recorded music in the future less and less, if at all.
Conclusion: The right to provide original music as training material for AI must of course remain in the sovereignty of the artists and rights holders. But this does not limit the results of a machine in the future at all, as long as available and above all growing knowledge from many legal sources with creative achievements of countless users leads to the same result. And Schniebert Hutzelman from Hintertupfingen, like millions of creative people, does not have to worry at all about the use of their original music for AI. Because the fact that the machine needs any of it at all for a third party is something that everyone still has to work out as an artist with their own audience and uniqueness … perhaps with the help of AI as a creative assistant. Maybe in the future it is an honor to be included in XY’s training set and it reduces the market value not to be included in it.