How much music creation and production is going to be replaced by AI?

Of course many AI tools will be used a lot to help in production, for example automatic generation of drum tracks to fit with composed music; automated EQing and mixing; automatic cleaning up of badly-recorded takes. This will make things quicker and easier for skilled and unskilled musicians and producers.

But on the whole, we don’t have a scarcity of supply. And most listeners are not constrained in their listening by any financial consideration, nor are most artists constrained in their amount of production by any financial consideration. Instead, the only scarcity is in listeners’ attention. We don’t have enough listeners, and they don’t have enough time, to listen to and appreciate the huge surplus of excellent music being produced all the time. (So, maybe AI could help musicians most by listening to music and generating realistic but positive comments!)

But that is treating all music as interchangeable. There is more to say if we break it down, and I see five main cases.

  1. There are artists where intimate, authentic details of the voice or playing are crucial. Examples: Bob Dylan, Stevie Ray Vaughan. AI imitation of that is probably not in the near future, but even if or when it becomes possible, it is not very valuable to listeners.

  2. There are artists where fashion and image are crucial, and the sheen of production reduces individual elements. Example: Backstreet Boys, Madonna. They could probably be imitated quite well, but it’s not clear whether fans would invest the same emotion.

  3. There are artists where technical virtuosity is crucial. Some instruments are easier to imitate than others for this, for example piano is much easier to imitate than violin or voice (even leaving lyrics out). Examples: Steve Vai, Yehudi Menuhin. But convincing simulations of virtuosity, even on violin, cannot be far away. But part of the attraction of virtuosity is admiration of the human skill and effort – just as in chess. The thrill of watching a virtuoso on stage is like the thrill of watching a tightrope walker. We wrote a little about this in “Should Music Interaction Be Easy?”.

  4. There are artists where ground-breaking novelty in style and composition is crucial. Examples: The Beatles, David Bowie. AI is not too clever at originating new genres, at least for now.

  5. There is music where the artist is quite anonymous, it is quite abstract and non-programmatic (in the sense that it doesn’t try to portray some real-world events), and we listen for mood, atmosphere, abstract pattern, often in a cerebral or solitary way. Examples: Philip Glass, Daft Punk (but see below). This is the most accessible to AI replacement.

Daft Punk are a pretty interesting case on the border. The style of a typical Daft Punk piece is a bit like (5) above, and could be imitated by AI. I bet an AI trained properly could us new, catchy, danceable pop songs like One More Time or Get Lucky. But Daft Punk innovated in at least four ways that AI could not easily replace:

  1. A human could ask an AI to imitate the Daft Punk style of sampling on Homework, but an AI could not, for now, originate the idea.
  2. They branched out significantly on Discovery, e.g. Aerodynamic and Veridis Quo.
  3. Some Daft Punk vocals and lyrics are robotic, highly processed, or pop-like, so the singer is quite anonymous; but consider the lyrics of Teachers, the sampled intro of Revolution 909 (with the police car), Giorgio by Moroder, and Touch.
  4. Perhaps Daft Punk’s greatest artistic statement – their farewell video, which is a movie rather than music, and which addresses issues of mortality and AI – could not be imitated by an AI music system.

Human artists can “jump out of the system” (Hofstadter). Ironically, Daft Punk pretended to be robots all along, but their ability to do this gave the game away.