Is Suno Sampling Copyrighted Producer Tags? (Audio Examples) 

@@RasmusSchultzthanks for sharing. I was thinking the same that now at this point since they've been exposed the only thing that can be done technically in the output validation stage is write conditional checks against the same database storing the audio and meta data used to train the models in the first place. if (sounds like Whitney Houston) { return "sorry can't expose a crime" } If they were ethical and smarter they would have prioritized implementing sound generation copyright protection safe guards from the start but greed and deception was more important evidently.

@RasmusSchultz so in other words it's like asking Bing or Chatpt to create a violent image. The system will refuse to accommodate the request. This is nothing knew. AI models have had restrictions on them pretty much out of the gate. Otherwise you'd have some pretty nefarious images floating around the internet. AI music is no different. The biggest thing here is that you couldn't reproduce this videos claim. Why? Because the results were "coaxed" using certain commands and tricks. Basically BS.

@@mephistro no, if you want a comparison, it's not like asking for a violent image - it's more like asking for an image of Iron Man and getting something extremely close to the actual Iron Man movie poster. The model contains a compressed representation of that exact image. There is an ongoing case over that as well.

It is not obvious. Record companies will either buy these companies or create their own similar AI music generators

Though you cannot e.g. use it in video without permission of the copyright owner AND it has to be a cover indeed, not an arrangement.

Found the one sane commenter

EXACTLY 💯 💯 💯 Suno has a pop up warning NOT to do that. This video is dishonest, he even covers up the propmts in the BOTTOM RIGHT CORNER of the screen so you can't see that's exactly what's happening. WHO PUTS AN INFO BOX IN THE BOTTOM RIGHT CORNER OF A VIDEO?? 😂 Like nobody would notice that. Spot on comment.

@@BerkmanHouse yes. User isn't directing the service to infringe intellectual property, but coaxing the service to reveal the intellectual property infringement.

At 3:33 it shows that the musical prompt was just "mellow trap", no? So I'd say Suno is infringing on rights, not Suno's users. Those terms of services seem to be just meant to cover their own ###. Also, regardless of what a user inputs, this type of AI tool should never be able to spit out exact producer tags (or melodies/lyrics/beats/voices, etc for that matter) as that clearly proves it was secretly trained on (huge) amounts of copyrighted music and this infringes on holders' rights on a massive scale. It would not know or be able to replicate/imitate these tags if it was not trained on cr music.

Ai does not sample. That's what sample-based software does.

@@scottbuffington5964 well it kinda does. AI certainly does not sit in a studio recreating, mixing and mastering every sound from scratch. Edit: I mean... did you even watch the first 25 seconds of this video 😂

It's not that easy to develop a generative music model. It would take years of specialization to learn how to do something on the level of Suno/Udio.

They will either purchase them or take them down and create new ones similar to them

@@sirhammon your brain is not an AI. Humans aren’t robots. Our minds don’t function the same way or produce the same outcomes as generative AI. And the two don’t carry the same legal protections. This AI is doing a spectral analysis and copying the same frequencies, as this demo shows. It’s not learning a new process to create a similar work. It’s plagiarizing the “tangible medium” of the recording.

A bunch of students 10 years ago generated every possible melody and copyrighted all of them. Its like musical library of babel. So technically they can sue anyone

@@andrewlloydpeterson won't hold up, as they generated them with automation, rather than with direct human creation. Copyright exists for *human* expression.

@@NathanaelRouillard what do you mean it wont hold up? They copyrighted it. Its done. They patented it. How can you prove something is created by a human or not? Music algorithmic generators existed since when mozart was alive before even computers existed

💯

@@mecd4167 email me at jesse@syncmymusic.com and I can make them available to you.

Nah

@@alexadigitalradio what claim did I make?

@@SyncMyMusic - The claim that the tags were created by Suno from copyrighted music they used for training is what I got out of the video. Perhaps that's not what was meant, but it can be interpreted that way for sure.

@@alexadigitalradio watch it again. I’m careful not to make definitive claims. This isn’t my first rodeo.

@@SyncMyMusic OK. Again, just trying to help as my day job does deal with countroom situations and evidence in law enforcement. A claim that isn't definitive is where it can become a problem if someone, perhaps not you, takes it to court. I'll bow out from here. Good luck down the road.

lol you really think DJ Khaled is uploading his tracks to Suno so they can plagiarize without attribution? If not, the machine is then copying without permission. Doesn’t make it legal just cuz some rando uploaded a superstar’s top single

@@YoPaulieMusic “to our knowledge” is def the issue at hand. If they’re not keeping an archive of the music/recordings, how is their model able to spit out what sounds like replicas of these producer tags? I def need to do more digging to understand these models better, but the AI companies claims that “no data is stored” doesn’t seem to ring as true considering what we showed in this video. And no “thank you” for revealing the prompts this time? 😎

It's really cool meta did get their licensing in order. Between both suno and Udio I have 100 unreleased tracks. I wonder how many users haven't released because of the copyright mess? 😮

precisely

I agree

Agreed. With such bogus terms its like The E3 hype game trailer is a metaphor for Silicon Valley right now.

@@tinagomez4418 swing and a miss. Wanna try again?

@@SyncMyMusic No thanks! with that response, I'm pretty sure I hit the target.

@@tinagomez4418 nobody believes you believe that

@@tinagomez4418 No, you didn't. Only in your head.

💯

Down to the waveform

@@davidcuny7002 it’s not the “steps to take” as much as the exact frequencies to replicate, as demos like this prove

@@bunnyyyyyy398 It's likely that the training data consisted of audio frames dividing spectral frequencies into Mel-sized bands rather using "exact" frequencies. By "steps to take", I mean the neural network can output frames containing similar spectral curves that approximate the training data. The point is that "samples" aren't stored in the network, but rather the ability to construct an output that approximates the training data. By way of analogy, it's akin to having the recipe for baking a cake rather than storing an exact copy of a cake.

You just described making a copy of an audio file...

@@bjornskivids Is there evidence to support this claim? Something can sound the same without being the same waveform. Even if the harmonic content is identical, if the phases of those harmonics differ, the waveform will look very different. The EnCodec audio compression codec Udio uses (at least with Bark) seems to have trained the discriminator to a resolution of a "64-bins mel-spectrogram", and I see nothing the paper where they attempt to maintain phase.

@@BerkmanHouse do you think the owners of the producer tags I highlighted in this video are going to take legal action against me or my source? Cause they’d be the only ones who could. I’d take my chances that they won’t be too mad at me.

Lolll weakest defense ever. “Don’t input the title of a real song in my song generating machine and it won’t violate copyright laws.” The whole point is that this reveals they are training on copyrighted material

@@SyncMyMusic I can only guess, but I'd guess no. I think your use of their property is fair and they don't appear damaged by it. It seems fair use if you are illustrating that the service was trained on copyrighted material (which I didn't think was really in question). My questions concern the nature of copyright infringement in these examples. If there's infringement in the output, but it's a result of infringement in the input (by the user); then surely the user is at fault and it's not fair to blame the service (Udio/Suno)? From Suno's view, the service is allowing use by the user, because the user agreed to their terms of use of that service. If the user then violates their agreement and uses the service in exactly the way they agreed not to do, it would seem reasonable for Suno to "be mad" at the user. (Especially if the felt their reputation/brand was being damaged as a result) To me it seems like: 1. User inputs copyrighted content, infringing on holder's rights and violating terms of service. 2. User "direct the Services to generate any Output that (i) infringes any intellectual property or other proprietary rights of any party", again violating terms of service. 3. User then suggests that the service (Suno) is responsible for the resulting infringement of copyright? If so, that doesn't seem a fair assessment (or fair use of the service).

@@BerkmanHouse they are liable if they don't remove that copyrighted data. If they remove it when made aware of the infringing data, then they are safe, but it seems they can't really remove it, so it is apparent that they are trying to hide the fact that it's in their dataset instead. They're still making use of the data though, just putting a system in place to try to hide their infringement.

Yes unless you download them

This was discovered before Suno started letting users upload their own audio.

The question that needs to be asked is why is it only @SyncMyMusic finding these materials. He's more than likely intentionally trying to sabotage Suno and Udio. Out of the thousands of people that utilise these platform he's the only one coming out with these results and claiming to have "sources"

@@tinagomez4418 ok I’ll help you out. I’ve been talking about AI on my channel for a year. A subscriber reached out to me with the Udio findings. That video led another subscriber to reach out to me with the Suno findings. When you talk about specific issues, the internet notices and interacts with you.

@@SyncMyMusic Help yourself and save your time explaining. That narrative doesn't prove my suspicion wrong.

I don't need you to believe me, but these were very easy to prompt. For the ones I personally promptef, they came out within the first try or 2.

@SyncMyMusic you're right, I don't believe you. I've been using Suno for months and know a lot of the little tricks to "coax" the algorithm into doing certain things. It will not create the producers label unless you type in the prompt to do so. But hey, why should anyone listen to some rando on RU-vid? Go try it yourself since it's free. See if you can reproduce this videos testimony.

"Bro, you just typed in the same lyrics." Yes, but he did not dictate the speech patterns (tone, pitch, cadence, etc.). He did not prompt the model to say the producer tag in exactly the same way the real producers do. The speech patterns had to come from somewhere.

@heypfist like I stated earlier, pitch, tone, inflection can be changed in other apps. Heck, right on YouCut, the app that most RU-vid videos are edited with, you can do all of that.

what if it's physical modelling?

Interestingly Google did pay for the music it trained it's AI music model with. Why do you think they would waste money on something not necessary?

​@@SyncMyMusic they'll have to answer to that, there could be a numerous reasons why they selected a different route with music in comparison to text. Maybe they wanted music that sounds in a specific way.

@@Vurt72 Meta also paid. So did Stable Audio. Sam Altman has stated that OpenAI won’t be doing music because of the legal issues. Draw your own conclusions.

@@SyncMyMusic again, i really don't care at all what happens to udio and whoever else making money of this. The tools they're using is what's interesting to me, and they're not going anywhere, they will only get better with time, like with all other AI. what i really want to do is train with my own audio / my own selection of it. Stable Audio sounds kind of garbage though, you can directly tell they have a very tiny library. But they'll get better with time i'm sure.

@@Vurt72 you mean you want to use these kinds of AI models consensually with your music? Cool, then we’re on the same team. 👍

Humans learn how to play guitar and learn songs they love from copyrighted music. Then they learn their own songs. You could argue that almost every song would have similarities to at least one other song. It’s no different than what AI is doing. The AI is still in progress and will learn to never reproduce exact samples or sections of the songs it’s trained on.

@@domehouse79 Writing original music does NOT distill down to just copying and pasting parts of other songwriter's songs into a new song and calling it yours. You said, "will learn to never reproduce exact samples..." ,errrm, where did you get that idea from? It's simply your opinion with no evidence.

do you think chatgpt and the numerous other text-AI's are trained on material with permission. They're not. Google, Microsoft, both gigantic companies are doing this.

Just like humans duh

​@@thaddeuscoreauhhh have you heard of Sampling? Its a thing that really exists where you take other people songs and transform them into new ones. No one calls that stealing

@@aaronseckmanmusicif you swap PCM with Dolby Atmos in your scenario, then yes there is. Also, that's not copyright, but a patent. Copyright is not the appropriate IP for that scenario.

@@SinfidelityMusic I’ve got a career making copyright protected music so I give a f@ck. How about you?

@@SyncMyMusic and did someone copy and release ur music? I doubt it

@@SyncMyMusic I can also DL and sample the shit outta anything. Doesn’t mean shit if i don’t release it or try make money from it.

@@SinfidelityMusic Thanks for your confession.

@@SinfidelityMusic my lawyers now know were to find you 🤠

This is interesting. Do you know if the visual artists are fighting back at all?

@@jmi_music Trying to! In the US there's a class action lawsuit that's going forward with multiple infringement claims (there's a website for it at "imagegeneratorlitigation" or see "Anderson vs Stability"). Last year we also funded a full-time lobbyist through the Concept Art Association. Though efforts this year there are stalled due to fundraising issues - a lot of people have already lost work and can't contribute anymore. There's also multiple European orgs lobbying, such as EGAIR. Seen Brazilian and other countries with movements a well. There's also now a new platform that artist have moved to due to scrapping concerns called "Cara". In the end, it's going to come down to regulation and lawsuit results. And the court results in visuals might set precedent for other industries as the proceedings started earlier.

@@jmi_music Sorry. RU-vid is eating my detailed response, but the short version is yes. Cases (which might set precedent for music, etc, as well), lobbying (regulations and opt-in approach), various orgs (US, Europe, Brazile, etc).

@@HolocoughSurvivor producers who pay for the licences to use those samples, complaining about being sampled without compensation. Fixed it for you

It's not illegal to use copyrighted music to 'train' AI. It's illegal to COPY or to BROADCAST music without permission. It's no different than a PERSON listening to a copyrighted work, and then making another work that sounds LIKE IT, but ISN'T it. Musicians have been 'learning' from other musicians and then using that knowledge to write other works for YEARS. It's called 'influence'. Musicians INFLUENCE other musicians. AI is simply being INFLUENCED by copyrighted works, it isn't 'distributing' them without permission. It isn't 'copying' them.

@@rockandroll798 I don't think this reflects the issue. As far as I know, a lot of the legal issues are about these AI companies training from copyrighted material without permission. If what you say is true, then Suno and Udio could just as well openly announce that they did this and there wouldn't really be any issue. I don't think either Suno or Udio have any interest or particular benefits from having their models being able to copy existing music. First of all, I don't think any users of these programs would find it very interesting to copy an already existing song, they want to create their own stuff, also Suno and Udio know that users can't use or release these songs if they are copies. The problem as I see it, is whether or not training an AI can be compared to that of a human doing it. And I think there are arguments for and against. For it not being an issue, is that these models still require human interaction and it somehow falls under fair use, since/if they don't directly store samples in the model, which I don't think they do if it works somewhat similarly to the image models. Jesse showed a lot of examples, but at least a good portion of them I think you would be able to make the argument that the user's intention was to get the AI to copy an artist, obviously it might happen by mistake, but at least from my experience, it is highly unlikely. Therefore I also think you could make the case that the users have some responsibility as well, just as you could copy a song on your piano, copy an image in Photoshop etc. Against it being ok, is that clearly, the CR material is integrated into the models and Suno and Udio have created some programs capable of copying CR materials. And even if that is not their intention, then they are somewhat responsible for making the CR materials available to be "copied/generated" without the right to do so. As these AI technologies get better, you will probably be able to generate songs pretty much on the fly, which means that you won't even need something like Spotify etc. You just tell the AI to create music in the style of whatever artists you want. Obviously, the music or any artist industry is not interested in feeding these AI companies if their intention is to grab all the money. So it depends on what is actually legal and what isn't, if it turns out that it is legal to train on CR materials, then it will most likely have a huge impact on the music/entertainment industry because people will copy things as they please or simply create their own stuff and companies will definitely start using this much more, they can save so much money and probably more importantly time. And if it turns out that it is not legal, I think it will kill most of these (US) AI companies and then people will turn to use free ones (which are completely without any limits) or maybe some foreign ones who don't care to much about CR.

The first one was found by accident and doesn't contain any word from the producer tag.

​@teko363 have you used it? All the first samples are from version 3.0. Suno is currently on 3.5. Also, not showing the whole lyrics on that first song. I found the first song. Dude only shared that one song and that one post on reddit and that's it.

No way this is a fluke. The timbre of the sounds is too similar to be a case of someone caps locking the words

People aren’t computers. Human creativity doesn’t work the way machines learn to plagiarize. Copyright law doesn’t exist to protect machines.

And how many people still use file sharing sites instead of just streaming/paying for their music now?

@@SyncMyMusic streaming sites are more convenient for people who don't won't to take the time to download mp3s and put it on a thumb drive. A Lot of people don't Even have a computer and only use their phones. Music consumption is also dying. Country music is the only music with its head barely above water right now. Background music will be AI generated and as the technologies get better the record labels will use it to put out new hits from old artists. Welcome to the New World Order.

"talk about"

Are you a software engineer ?

@@justin.johnson Nope, just very interested in AI for a long time and have been using LLMs, image and video creation, and audio generators for a few years now. I wanted to understand the technology so have been delving into the research papers and using the tools. I started using the Open AI sandbox back in 2021 and got hooked on the technology

It’s effectively spectral sampling. If you know how that works. The tags are being regenerated. Suno can’t hide behind “oh some we let some rando upload DJ Mustard track so we can do whatever we want with it “

@@bunnyyyyyy398 Producing a spectral sample is entirely different from creating an actual sample. Suno does not simply sample music and insert those samples into tracks. My point is, Suno would not replicate a producer's tag so precisely and then randomly place it into a song. In my opinion, these cases involve users taking samples from songs and incorporating them into their tracks using the upload audio function. It's peculiar that this is occurring now, just after the upload function has been released. We know a lot of people hate Suno and Udio, and we have to consider if they are doing it to discredit them.

@@newsjamgroove0 Nope, we didn't upload any audio to create these outputs (only prompted them). And this was discovered before Suno released their audio upload function.

​@@SyncMyMusic Yeah we'll just take your word for it... 😂😂😂

You can see the prompt for that first track with the "CashMoneyAP" tag included in it. There's no prompt for those words anywhere.

​@@SyncMyMusic Another lie. It's the title of the track. 😂 Suno reads that too. Dude you're so dishonest.

If Suno would just say what music they trained on, there wouldn’t be this ambiguity.

@@jmi_music I generally agree with this position, but also, legally, they do have the presumption of Innocence and a right to guard their trade secrets in a highly competitive and new arena. They don't owe us disclosure by default, and opacity should not be used as evidence of guilt. If parties suspect that Suno or other AI are violating copyright, the burden of proof is still with the accuser. Do they have enough evidence to make it to discovery? It will be interesting to see. This is definitely a fascinating and important case. Thank you for engaging. Cheers!

No. Human brains are not generative AI models. They function entirely differently for different reasons with different outcomes.

@@bunnyyyyyy398 i didn't make the claim that the human brain and AI are the same. Don't forget we still don't actually know what these sites are doing, ripping the music and using stems or creating the sound itself with physical modelling synthesis. I'm asking if these websites are listening to all music, and then reproducing something like what they have heard through Physical modelling, is it actually copyright infringement? i don't think it is

@@BuddaB911 Why no AI fanboy gets that it's not the AI model on trial, but the CEOs of Suno and Udio? They illegally downloaded copyrighted music to train AI. Whatever training of AI has to create a local copy on their servers, so to go out from Spotify, RU-vid etc, which is forbidden. The CEOs don't listen to the music, they aren't artists, yet they profit from artists' work. They just violate copyright in the very moment they download the music, which is supposed only to be listen to.

@@federicoaschieri My friend you are way out in space with this comment. how do you even know if any music was downloaded illegally? What if they bought the Vinyl/CD? what if they recorded it themselves? what if they have a system where the system listens to the source no matter where it's from directly? Do you know the CEO's have never made art or music? Don't we all who make art profit from engaging all our lives in other artist work for inspiration? Non of us have answers to most of these questions therefore you can't make the claim that anyone "Copyright" was violated until you know for sure if it's a copy or synthesis

@@BuddaB911 Transferring the tracks on the computer is always illegal, it doesn't matter if from Spotify, RU-vid or a CD. Copyright, as you appear not to now, prevents unauthorized copies to be made. Whatever system they use must do that before training. Your law-fantasy arguments need no answers, because are illogical. As I said, *we* listen to music, we don't download it. Law says that the first is allowed, because it doesn't make perfect copies of the digital files, while the latter does. So those pirate CEOs violate copyright.

@@MRNoory-vt3nb this is easy to say when you haven’t made your living using the copyright system.