I found an awesome 80s style cover of one of the Kpop Demon Hunters songs on YouTube. I still like it — but the channel has dozens of videos including 5-6 of any given song, they make more than one a day, and when people ask them to go on Spotify, they say they won’t — which is interesting because Spotify just banned AI generated music. So I’m 99% sure they’re an AI band. My point is, I can’t help liking the song, so I feel like we’re screwed because that could happen to anyone. (FWIW I’ve downloaded the song so I can play it offline. So they don’t make money from my plays.)
So my question is. How do we know? And what can we do?
You must log in or # to comment.
I always find the vocals to sound scratchy, like a lot of AI voices. Infoections are often random or unnatural. Also, the music usually sounds incredibly “average” or generic. There’s usually not a specific tell. It’s one of those judgement sort of things.
Often when Instruments Pause their effects wash over the vocals. Especially with guitar distortion.
If you suspect it’s AI, it probably is. Churning out large amounts of songs, no face to associate with the music, production quality that is too perfect, and so on…
That’s why this is so rough, at the very barest minimum anything AI generated should be labeled as such, but this is the world we’re living with right now. Support your local artists!
You goddamn well better not listen to something you like if there’s even the tiniest possibility that AI was involved
Hella replies. Zero addressing the point.
For sure? I don’t think it’s likely there’ll be a way. You can only get best guesses.
In my case it’s mostly intuitive, a sort of uncanny valley but for listening that clues me in that something’s off, and then I start looking for stuff actively. If the “band” has many different genres or varied styles, if I go to older videos and the singing is different or the voice is altogether a completely different one, if there’s no social media presence, if there’s no interaction with the comments, if the account’s relatively new, if there’s no sense of improvement in a natural human way (as in they clearly got better at playing), etc.
But nothing will ever be enough for me to be 100% sure unless it’s extremely evident.
Personally I get my music from the past, or from countries that haven’t swallowed the LLMbecile Kool-Aid yet.
https://youtu.be/3Nlb-m_vKYM?t=2m57s
tldw: ai generated music is very good at making music that sounds physically correct to a human ear, but since it’s all generated as one track, it doesn’t have any source stems. and as a bonus it’s probably been trained on shitty mp3s!
so:
-
download the song, somehow.
-
use an (ironically ai) stemmer to split the song. the latest audacity releases come with this tool preinstalled under the openvino effects.
-
then listen to the stems individually to listen for mush. if it sounds noisy with some bleeding between each stem, and/or if you can’t hear the rests where an instrument isn’t playing, it may be ai.
effectiveness may vary by genre.
I actually have the latest Audacity, for macOS if that matters. I don’t see anything called openvino in the Effects menu.
I’ve only really ever used Audacity to shorten songs, like to make ringtones.
-
Only listen to artists who you’ve seen live?
Edit: I feel like this could be seen as dismissive, but truly the solution is to embed yourself in your local music scene. Support real artists who are people in your communities. Dont let algorithms designed to benefit big tech/record labels determine your listening habits. Keep an eye on the schedules from your local venues big and small, or follow their RSS feeds. Listen to community radio if you can. Follow people from your local/DIY music scene on social media. We need to build community around music again and reclaim a culture that supports artist rather than content creators and isnt controlled by large platforms ran for profit. Great music has existed on these fringes literally forever and still thrives in smaller communities today.
For further reading: “Mood Machine - The Rise of Spotify and the Cost of the Perfect Playlist” by Liz Pelly
Depending on your music taste, this is bad or even impossible advice. A significant number or genres I listen to generally don’t perform live, let alone anywhere near me.
Yeah, I’d love to see SrenHlimMrews live, but … I don’t live in Chengdu where I could haunt underground clubs for the rare times they pop up to play.
What I described is more of a necessary culture shift to disincentivize “AI music” as digital content, which is unavoidably pervasive in the muddied waters of cloud based algorithmic music streaming. The platforms where these things propagate are built explicitly to enrich those who already control the music industry, and pro-artist people-centric communities where oxygen isnt given to AI slop must be decoupled from these platforms.
As you say, diving into the local music scene isnt an answer for everyone to find the music they love without risk of slop being mixed in, but I do see it as essential to following the path toward revitalizing the cultural importance (and livelihoods) of artists over content creators, engagement metrics, and lopsided payment structures that have allowed AI slop to pollute our lives.
There is no reliable method and it will become harder to tell as the technology improves.
That’s why people complaining about “AI slop” is somewhat weird and dishonest. Sure, there is lots of terrible generated content, but it’s increasingly undetectable. People are just complaining about the cheap stuff.
It’s not weird or dishonest to complain about AI “art”. I want actual human art that has actual meaning or effort put into it, because that’s important to me.
Complaining that AI “art” is bad is one thing, but disliking being tricked by some soulless inauthentic fake “music”, that only sounds real because of how much actual music it has stolen is another thing - and is also a totally reasonable thing to complain about.
I don’t think there’s anything weird or dishonest about wanting human art, personally.
Me to a vegetarian: “Ah, you can’t even tell there’s pepperoni in there, can you?”
upload speed is the largest indicator. No human publishes more than 1 song per day. In fact I’d go as far to say that more than 1 song per week at the minimum is sus. Music, like many other things, is hard work.
Sometimes it can be hard to tell, but usually it’s the lyrics that give it away. The old Sora models made very basic lyrics that all had very simple rhyming schemes. The newer models tend to be a little more adventurous. Another way to tell is if the vocals sound unnatural in how a human would pronounce things or the cadence is unnatural.
That said, I don’t think there’s anything wrong with liking an AI song. Not giving them the views is good, since AI music is exploitative, but if you like the song, enjoy it. I’ve got a few that I like to listen to. The way I view it, as long as you’re not giving these companies money for producing or consuming AI slop, you’re not doing anything wrong.
I agree with you about the lyrics. I think if it’s a cover song, it’s gonna be harder to tell, in a vacuum. I suppose I have my answer in how many songs this channel is churning out. What really did it for me was where the quality would dip. They act like they’re actually doing these as performances in a studio, but there are lines being delivered that no reasonable person would sing that way and stamp with their approval. Because with a cover song, the lyrics have already been written and you have the song, you just change it in some way.
There’s no reliable method of detection, you can only really listen for clues.
