Critical Response #1 |
Februrary 4th, 2023
Music 356 / CS470, Stanford University
Critical Resonse #1 - What do you (really) want from AI music generation
I think MusicLM is awesome. It presents a really cool way for people to create the music that they imagine (at least to an extent). From what I understand the algorithm uses a labeled database of samples, which on a larger scale seems really unwieldy to create. Fundamentally I think this kind of generative tool can help people not trained in art, to make their own. In today’s world many are forced into employment that they may not enjoy, and learning something like a musical instrument and music theory is a massive privilege that many don’t have the chance to explore, especially later in their lives. It’s a profoundly human desire to want to create, and for many they can’t afford to create the things their mind may want to. I think many in artistic communities can be incredibly snobby about inexperienced people wanting to create, saying things along the lines of that these people haven’t put in the work to make music or to paint. I think giving people tools that allow them to concretely be creative with little training is amazing, cause it really helps those who are creative but didn’t have the resources to learn.
Something I’ve seen echoed a bit on the internet is the idea that because the person generating the art (be it music or an image or anything else) isn’t spending actual time doing it, it’s somehow invalid, and they must just be “too lazy” to learn how to do it properly. This totally ignores the fact that many in today’s world HAVE to work incredibly hard in incredibly boring menial jobs, and want some way to be creative in their off time without having to cram in even more work into their already exhausting days. One complaint though from these same people complaining about the laziness thing, is the fact that these algorithms often are training on people’s art without their consent. This is obviously very evident in the case of MusicLM where they immediately got hit with copyright issues.
This really asks an interesting question in my mind, why is this different from a human looking at reference images, or getting stylistic inspiration from other sources? I believe that it is somehow, fundamentally different but I really cannot begin to explain why. One would never tell an artist that in order for their art to be legal or legitimate they must never see art in their lives ahead of time, that would be absolutely inane and ridiculous–so why do people expect these algorithms to generate completely from scratch? It also brings up the question of how much would you have to rearrange a wav file for it to be a new song or a new work? Certainly our feature artist is still very much the original song, but at what window size would this change? Surely if we went single sample by single sample and re-arranged them we could make it sound completely different, and there would be no way to reverse engineer the old song. If we created a new sound from scratch which just happened to be an “anagram of samples” of some other song (ie same exact set of samples in some other order), nobody would question it’s legitimacy, but if we started with one of the files and re-arranged it into the other, would this all of a sudden be an issue. It feels like some sort of unenforceable rule, is there some length of time where the audio chunks wouldn’t be recognizable? Would that even be the line?
Sorry for this very rambly response:
TLDR: Ai art allows those for whom artistic training is inaccessible to be creative, and at what point are two audio files no longer just mixes of eachother.