The problem is that YouTube AI thinks your recording is the same as every other recording, because it doesn't understand the difference between composition and recording.
When real songs and AI songs can't be told apart even by groups of fans working together, the artists should be scared.
If AI can be so convincing, then your publisher won't we wanting to pay you to make any new songs if they can just pay an AI guy for a few days to crank out a new song.
> So probably what's happening (some-how, some-way) is that the ID mechanism is trained to find songs that are "the same" and in pop music they're only "the same" if they're the _same song_.
Right. Pop music has a lot of variation in timbre even between 'covers' of the same composition, so if the notes match and the overall timbre matches it's a safe bet that you're looking at two instances of the exact same recording, or at most two versions by the same performer. In classical that just isn't true, the timbre of acoustic instruments is basically fixed and the notes are defined by the original composition. There's just too little variation for an auto-id system to work with.
I have gained the impression that YouTube has two distinct techniques within Content ID, one for claiming melodies, and another for claiming specific recordings.
I’m absolutely confident that this is doing melody matching: these recordings are trivially distinguishable from any professional performance, with completely different instruments and playing styles, and with much lower quality singing.
For example, the most repeated claim was by “AdRev Publishing”, claiming “Crimond (The Lord's My Shepherd) - FirstCom” three times. (That song was also claimed once by “Capitol CMG Publishing and Adorando Brazil” as “The Lord's My Shepherd I'll Not Want”.) The first two times, I was accompanying with a piano-sound keyboard in the traditional four-part harmony—admittedly they sound fairly similar to one another; but in the third, an Indian was playing, using a piano-and-strings sound on a different keyboard, using Indian harmonisation (which is quite different).
I think we even had an unaccompanied song claimed once, matching throwaway0b1’s report.
These are completely different performances from whatever recordings the liars may have provided to the Content ID system. The melody is the only thing they will have in common.
I said that two actual and copyrightable performances of the same melody are arguably indistinguishable.
Edit: How would an automated system know, that you did not just non-transformatively alter another person's performance, instead of performing yourself?
The problem isn’t what they store, but what they produce. a « database of songs » is a problem only if gives you access to the copyrighted content. Wikipedia contains the key characteristics of millions of songs (aka their names and author) and yet nobody complains about it.
Those AI produce original content. The means by which they produce it shouldn’t be relevant.
I once asked Android Auto to play "act 2 from carmen" and it responded with an absolutely bizarre rendition of the Toreador Song that had been programmed into a MIDI in the voice of birds chirping. Just crazy. Sometimes I think the algorithm they use at YouTube Music optimizes for paying the least royalties.
If no one told me they were AI, I’d probably assume it was a parody group or a house band messing around in the studio. It doesn’t sound like an artist writing with intention.
And I’d wonder why they encoded at 32kbps with a RealMedia codec from 1998.
Has anyone actually verified that the song was generated by AI? It should be trivial for a good singer to mimic the voice and style of a famous artist and claim it is AI as a way to get some notoriety.
I think this is an S+ tier example of why you shouldn't pay for ChatGPT.
I just tried for a few minutes to try and have it "help me remember song lyrics" and where it is somewhat correctly guessing the songs, it's disheartening to watch it absolutely butcher it. I can see they are trying to appeal to copyright but this is not how you do it.
If I was a CEO of a company and I was considering paying money for ChatGPT this would be the demo that convinces me not to. What use is an AI to me that can't even get song lyrics that I can find literally anywhere else on the web? The only thing you're doing in my eyes is making an absolute fool of yourself showing just how much your willing to break your own product to the benefit of what?
It's a fool's errand to intentionally hobble your own AI, if anything you're exactly making the case why we SHOULDN'T use your AI and instead use anything else that will be unrestricted. AI is being killed in the cradle by copyright.
I'm not worried about AI taking over if this is going to be the result.
I agree with you. When gwern investigated AI folk music in 2019, I realized it could generate a wonderful variety of music, full of soul. Be sure to listen to several tracks before making up your mind. My favorite is “crossing the channel”, since I think GPT made a mistake at the beginning, and then generated the most reasonable sounding not-mistake, which turned out to sound so cool.
My goal was strong, memorable melodies. Star Wars, not Marvel. GPT can come surprisingly close, if the input data format is right. Unfortunately I don’t think anyone except gwern has noticed that the input format is crucial: https://gwern.net/gpt-2-music
Yeah, you’re right. I think this is a case of simplification in communication over perfect technical accuracy. What he meant was more that song “similarity” (as in artist, album, etc) is not taken into account by the algorithm. But yes you’re absolutely right.
However autotune has existed for decades. Would it have been better if artists were required to label when they used autotune to correct their singing? I say yes but reasonable people can disagree!
I wonder if we are going to settle on an AI regime where it’s OK to use AI to deceptively make someone seem “better” but not to deceptively make someone seem “worse.” We are entering a wild decade.
Someone had to take the time to perform the song, there's no reason they wouldn't take a few minutes to change/improve the lyrics, even if an AI produced it.
I suspect it is not AI produced and simply a simple song.
"I think that is part of what is making it difficult for the untrained ear to differentiate between these AI-generated and non-AI generated tunes," music journalist Hattie Lindert told CBS News on Tuesday. "It's pretty convincing when there are so many Drake tracks that AI can train from."
So I generated a few songs using the same template, and while the songs sound OKish - each next one generated from the same template sounded VERY similiar - some sonics and sequences were the same, but they sound like nearly the same songs. So, while it looks like it is making unique music per each generation - the uniqueness of the music is likely made to escape being fingerprinted as the same song. A person, listening to all of the songs that I made, would likely say - "Didn't I just hear that song?" After hearing the first one. It's not like it's the same genre, it's a pitch shift here a filter cut off there, a sequence randomize there - and poof - new song.
---
What can the artist do if they don’t have the money to pay for a sample?If they don’t have the money, they need to replay it or re-sing it and that eliminates two thirds of your cost. So that’s a big saving.
lol or you can run it through AI programming and reshuffle, reroll etc. paying for samples is going to go away.
There are undoubtably a ton of variations of those chants over the past 120 years and probably millions of recorded performances of those chants. Any of which would be subject to a new copyright on the date of that performance or rearrangement, regardless of the original date of the composition.
A work being widely performed is a common situation that complicates attempts to solve the problem automatically. An infinite number of monkeys randomly singing a song will eventually sing one that sounds exactly like yours. How does an algorithm determine the difference in intent between my randomly singing choir of monkeys and your church's performance? It can't, as it compares audio -- it doesn't reason like humans do. Content ID doesn't know who your performers are, who recorded it, which sheet music your performers were reading from, where they made mistakes, and/or what licenses you have. (and FWIW, any given human watching the video wouldn't know much of this either)
Two audibly identical works can be non-infringing, and two audibly differentiable works can be infringing. ContentID is an approximation -- it does not and never will have enough information to accurately judge everything.
reply