Menu
What's new
By:
Filters

Can we sabotage music-scraping AI?

Technically there is “poison” in virtually any data, part of developing a reliable AI model is training it through the junk, GIGO avoidance has been part of this technology for years
 
Very interesting thread (and idea)...
I would also be very excited to know of a way to poison the AI Datas...
Of course, only as an academic subject of interest...
 
Yeah I don't see a way to do it with music, unless we all release lots and lots of tracks that are incorrectly classified by genre.
 
aeliron said:
Yeah I don't see a way to do it with music, unless we all release lots and lots of tracks that are incorrectly classified by genre.
Click to expand...
Earlier this year someone on VI-Control made a witty comment about how all of us composers should flood the internet with tracks made from nothing but ukuleles, xylophones, and hand claps. Whether AI is looking for whimsical, epic, ethereal, or dystopian, everything it spits out will sound like a commercial for Puppy Chow.
 
Polkasound said:
Earlier this year someone on VI-Control made a witty comment about how all of us composers should flood the internet with tracks made from nothing but ukuleles, xylophones, and hand claps. Whether AI is looking for whimsical, epic, ethereal, or dystopian, everything it spits out will sound like a commercial for Puppy Chow.
Click to expand...
The joke will be on those composers. Puppy Chow will license those songs, sell more product than ever, and the AI company will take away all those sync fees and royalties from the composers. Best to focus on making better business decisions instead of farting around with AI games. 🫢
 
There's some truth to it. I honestly think that the reason why AI music has progressed so poorly are two fold: 1) The number of good tunes vs. the amount of crap on the internet including most music library sites is about 1:100 respectively. 2) The amount of erroneous and overly complicated meta data attached to said tracks is hard for any intelligence to figure out because a lot of it is just designed for search engine placement rather than honest description of the tracks.

Where it is easy to feed AI a picture of "woman" until it figures it out. It is relatively hard for AI to figure out a track description like" a little sad, but not TOO sad, funk with a little blues influence, influenced by Jimi Hendrix, ACDC, John Williams and Mozart. Nudity."

I honestly think it would be pretty easy to train AI and I have an idea how, but so far the human influence in training AI know little about music if anything at all. So, I have half a mind to contact some of these companies trying to train AI to do music, get a job for a year "teaching" it only to sabotage the project with false data and get fired only after they paid me a hefty sum already.
 
I haven't had time to read the article, but a thought I've been toying with for awhile is, what would happen if we feed AI it's own output back into it, let it train on its own garbage, and just keep doing that? That would effectively make it an inbred moron, probably in a matter of minutes, spitting out musically incomprehensible garbage.

The difficult part is how to actually accomplish that in real life. I mean, I'm sure Epidemic Sound is already training some AI on their catalogue - so how do we get the AI garbage into that?

EDIT: @José Herring This is your mission, should you chose to accept it...
 
thesteelydane said:
I haven't had time to read the article, but a thought I've been toying with for awhile is, what would happen if we feed AI it's own output back into it, let it train on its own garbage, and just keep doing that? That would effectively make it an inbred moron, probably in a matter of minutes, spitting out musically incomprehensible garbage.

The difficult part is how to actually accomplish that in real life. I mean, I'm sure Epidemic Sound is already training some AI on their catalogue - so how do we get the AI garbage into that?

EDIT: @José Herring This is your mission, should you chose to accept it...
Click to expand...

With image AIs they have watermarks embedded to prevent that. Audio AIs likely will have that as well. When you run it locally it's trivial to just remove the step that adds the watermark as far as I know, but you'd have to literally flood all places where music is with shitty AI music, and even then the people handling the data mining for training would find a way around it, like pre-selecting the music more by it's popularity among real listeners.

I think the reality is that we can't do anything meaningful in sabotaging their operation from the outside, we need to do lobbying on the legal side to get this shit banned. It's the only way.

The concept art association is doing that already for artists:
www.gofundme.com

Protecting Artists from AI Technologies, organized by Concept Art Association

We cannot believe how quickly this year has gone and we are getting c… Concept Art Association needs your support for Protecting Artists from AI Technologies
www.gofundme.com

@José Herring: did you ever reach out to them to talk about joining forces?
 
I still love the fact that when some guys used AI to generate a song that sounded like 'Drake and The Weekend" that nobody could tell the difference between it and the real thing.

The utterly lame 'music' that is put out by Drake and his 'producers' might as well have been put together by AI in the first place.

I think Drake and his crew are a perfect example of someone being a victim (or unknowing contributor) of the mindset of 'the glorification of mediocrity' that has crept into society over the last 10 years or so.

You know...the mindset of 'let's do as little as possible and see how much money we can make off of this and then repeat ourselves over and over again'.

Point is...let's hope for a better future for our children than for them having them listen to a bunch of soul-less dribble that is foisted upon the masses.

I really hope that somewhere, especially in the commercial fields of music and art, there are some 'AI genius's ' who might actually have some soul or an ounce of humanity left in them and aren't just a bunch of 'corporate lackies'.

Is that too much to wish for for the New Year and beyond?
 
thesteelydane said:
...what would happen if we feed AI its own output back into it, let it train on its own garbage, and just keep doing that? That would effectively make it an inbred moron, probably in a matter of minutes, spitting out musically incomprehensible garbage.
Click to expand...
Actually, that's pretty much how they trained AlphaGo and all its spinoffs. By setting up the right sort of feedback loop, the AI was able to improve on its original training data and come up with novel ideas. I'm equally excited and terrified by this sort of thing.

Of course, that was in a much more narrow domain. There's a few years still before anything like that happens for languages, music or even images.

Another thought is that as well as defending current copyrights and trademarks, it's time to rethink copyright and IP law from the ground up. Humans learn the same way as the current generation of AI: you look at examples of the thing you're trying to learn, and whatever you produce will inevitably show traces of your training. There's a spectrum from "it's transformed so much it looks completely original" though "homage to..." down to straight out plagiarism. And there's not really a clear line. But Bach and Vivaldi would have struggled in today's legislative environment.
 
aeliron said:
Interesting article on how creators are “poisoning” AI scrapers. Wondering how this can be done for music, if at all. I mean, just as a theoretical exercise, of course. (Coughs)

theconversation.com

Data poisoning: how artists are sabotaging AI to take revenge on image generators

As AI developers indiscriminately suck up online content to train their models, artists are seeking ways to fight back.
theconversation.com theconversation.com
Click to expand...
My partner works alongside Daniel (part of ADNS) - and he's a super interesting thinker. Thank goodness places like ADNS exist.
There's loads of interesting reading in this space - and projects - for those interested. Most exist in the academic space right now, but sometimes leak into more public forums.

ADNS is such an important research center right now in Australia (alongside the school of Cybernetics at ANU). Ditto places like Arts Law in australia (shout out to the new CEO Dr Lousie Buckingham - awesome choice and fab human being)
 
You must log in or register to reply here.
Top Bottom