Is Udio Reproducing Copyrighted Songs? (Audio Examples) 

Thanks for watching all! Before you ask:"Can you share the prompting method please?!" just know that we won't be publicly sharing the method that allowed Udio to generate these tracks as we don't want to help Udio clean up their mess. That's their job and responsibility. Feel free to play around with their model for yourself and see what you can come up with!

The best way to use udio,is to type your own lyrics, and add an acapella tag so u get vocals that sound original then u can just make a beat around those vocals

A voice cannot be copyrighted. According to the legal decision in Midler v. Ford Motor Co.: “A voice is as distinctive and personal as a face. The human voice is one of the most palpable ways identity is manifested.” This ruling did not impact general copyright, but its subsequent legal interpretations means that while a recording of a voice may be copyrighted, a voice itself may not. As for the lyrics. Udio did not just create those lyrics. Whoever prompted the song just included them in the prompt.

I love the optimism of the thought "I wonder *if* they trained on copyrighted music, and *if* they licensed it" Of course they did train on copyrighted material, of course they didn't license it.

Excellent work, Jesse. Thank you for the reframe especially.

Recently I heard someone say, "If these platforms don't let you use artists' names, it's not because they're good people, but it's because they want to prevent you from finding out what data they've trained their models with." 😉

Awesome video, @SyncMyMusic Prompt: "Create a song that explains how Udio is about to be sued by every musical copyright holder." Udio: 🎶"Sue, sue, sudio!" 🎶

This is bizarre. I just posted a link to an example of Beatles song generation, and the video is now modified and the Beatles likeness and Beatles lyrics has been removed, but the video URL is still the same. That is one of the strangest things I've ever seen. It was "Chase The Sun" with both music and lyrics almost identical to Follow The Sun, except now it's suddenly a different song.

I remember when the big 4 went as far as arresting people for sharing crap quality mp3s online.

Thank you for this, shared to all my social medias ♥

Thanks Jesse!! Awesome Video 💪💪

The last phrase was very positive. Congratulations for being able to see that bright side despite everthing's happening. I feel that... Maybe... People will claim for more and more organic and custom material. Considering that AI is a soul-less music provider. Is too soon to know what will happen... Meanwhile, I'm backing to my roots, playing my old friend, the acoustic guitar. Which I always used as a therapy. (Btw, that's a great perspective to see the process of making music, as a therapy). I wish you all success. Greetings from Brazil.

You have asked some interesting questions!...good job.

Its NOTillegal. Because its the same as a cover bands rights to recreate music. Its a recreation and not an actual sample from the song.

GREAT vid!!! Fascinating and mind-blowing Beatles renders

I prompted 90’s, New York, boom bap and it gave me vocals that had the exact tone, cadence, flow and inflections of Fat Joe, another sounded like Ghost-face (but a bit more like Action Bronson), and one that was for sure Raekwon. I don’t know how it could be that similar without the models training in their catalogs.

I just tried creating a song by copying existing lyrics from a semi-popular song (57 million youtube views) into the custom lyrics field. It's definitely copyrighted, published in 2015 and is published by the productionmusic label ExtremeMusic. They are well established and successful. Since productiomusic would be under threat too I doubt they would let AI train on their music. I entered just about 20% of the overall lyrics. When the songs were about to be generated I got a "moderation error" saying "Your credits for this generation have been refunded. Please try modifying your prompt or lyrics and try again." My prompt was just "neo-classical epic emotional" So that can't be the issue. It's insane but this suggests to me that they have stored those lyrics in some database and stopped the creation of the song as the algorhythm recognized them. But why would they store those lyrics? Just because they care deeply about copyright and don't want anyone using lyrics from an existing song? Did they build a dedicated system that randomly stored a ton of lyrics from existing songs, even less "famous" ones? Or... did they store the lyrics in a system that is used by AI to generate songs and they just try to prevent anyone from finding out with this trick?

Thanks bro. This is needed.

Hi Jamie just joining the channel. You're absolutely right about us humans continuing to create music.. I've been working on a solution for years and we're almost there

thank you so much for this

Nice quote at the end there, thanks.

If they didn't just outright train it on copyright music... How much do you want to bet that they trained it on music from youtube? Probably picked up a lot of covers of these songs as well as original tracks that escaped the content blocking, even if they excluded the official ones. That wasn't John, it was PAUL on both the first 2! Close though on the first one, more raspy Paul though (think Helter Skelter).

Is there a way to prompt the chord structure. Each of these always generate two songs, but none of my udio or Suno sound alike.. what can I do to control the chorus progression?

“Cleverly Prompted” bro they typed the lyrics in obviously 💀💀💀

Haha even Shazam detected that as the originals!

There will be a renaissance of original music that is authentic. For those of us that are truly musically inclined, there will be a future. Focus on finding you own original sound.

Personally I would like to believe AI music will allow some talented people to break into a music career where normally they wouldn’t have had a chance. While this is also opening the floodgates to a bunch of noise, there are gonna be those songs that just rise up.

In udio you're not allowed to upload lyrics if you don't own the copyright, so that's what happens if you do it

Great video!!!

Nice outro conclusion!!!👍🎶

Really well done, Jesse. I've been experimenting with both Udio and Suno and admittedly blown away by the results. That said, I really wonder how they "gather" the material to train their respective models. Chat GPT, Dall-E etc. are all facing multiple lawsuits for "scraping" copyrighted material from multiple sources. As a musician, I wonder if these AI generated songs will make it into streaming platforms, commercial radio, Billboard chart etc. and if so, are they flagged or watermarked as being AI generated? (suspect most people won't care - especially if it's a good song). ...Interesting times!

Yeah, I discovered fairly early on that melodies and chord structures of at least Christmas music seemed to hold pretty strong coherence. I generated an orchestral piece of Hark the Herald Angels Sing. I didn't think to check copyrighted music though. I was just surprised to hear the A.I. just straight up generate the actual melody and chords. I figured Udio knew it was was public domain so they allowed it to happen. Just my own speculation, but I'm wondering if they were working on a way to do generative cover songs with a branch of their model that retained melody, lyrics and chord structure and they were experimenting with Christmas music, so they filtered the music they were scraping from which had key words like "Christmas" or "Festive" through this 'cover music' variation of their A. I. model. But IDK how it all works. That's total speculation.

Great video Jesse. Theres no denying now where they get their training from. upt o courts now I guess

Love the reframe 💜

A very good and thought-provoking video. These are very interesting times for all in the music industry, and my personal thoughts are that artists who produce very 'same-sounding' work, will feel most under threat. So those who push the envelope and continue to evolve their sound, will stay ahead of the 'learning'. It will push us to take more risks as you say, and for what can be a very safe and static landscape, I think that can only be a good thing, right? Thanks!

I think Udio should have the right to let the AI listen to lots and any music and learn, just like humans do. I do not , if I compose a song, try to find , who I have been influenced by and then ask their permission to use a short sequence of chords or feel, or style and ask how much I should pay to use that idea or chord progression or for example ABBA -like mixing method or Hollywood Strings sound. We all learn everything we know from the world around us. There are laws that protect our copyrights and those laws define, how much similarity a panting, a book or or a score can have without it infringing that copyrighted or creative work. God luck, Udio! You help us generate better music and challenge our creative minds to become even better with what we do as humans do.

One thing that no AI system can do, is replace human live performance

I have a question, how it generates those instruments? How does it replicate Les Paul exact sound, Juno, Jupiters, CS-80s, etc? Big brands like Arturia take years to model any synth yet AI creates it in 3 min with no sources, I feel what it does it steals stems from the songs it has been fed and transpose them or even modify the melody like you can do with melodine.

Great insight ' thanks

Hey Jesse, I've been a rebel since the beginning of my intro to sync licensing. I rarely used reference tracks and wrote what I liked to write. so, after almost 10 years struggling with this a lot of my music is getting the attention of a few libraries that didn't want anything a few years ago. I said from the very start the old school music will be back because it is good and not like the junk that is out there today. AI can kiss my behind since you have to pick your battles and fight them as they come. My strategy is not changing. That's my 2 cents, take it for what its worth.

Reminds me of the early days of image generation. If you prompted them in just the right way, you could get them to sorta recreate certain very very famous images. But as the training data expanded and diversified, that problem went away.

What do you say to a person that says they don't have a problem with this any more than they do with you covering a favorite song? The violation is in the USE of the generative production, not in the production itself. So generating it, shouldn't be illegal. But if I then used it in my film or commercial, just like if I used your cover of copyright material in my film or commercial, THEN i would be in violation and subject to litigation. I'm just asking.

Honestly, it's kinda cool that your friend was able to do this. I've tried to do this on udio. It's difficult to impossible to figure out. Also, I'm really curious how your friend got around the moderation system to make this happen. You're making it sound like the standard use case, when what you're clearly talking about is a hack. Please share the details of this exploit. Thanks.

The more videos I see like this the more excited I am about the type of music that I care about having a renaissance. Unless we’re willing to allow culture to die, creators will keep creating and making and combining new sounds. Thanks Jesse for leading the charge.

No doubt copyright problem are coming soon but! I think IA open for us the opporunity to personalize our music, new sounds, real emotions and being original. This bring us to a new way as a composer without having to using IA.

I did the same thing. Asked for a crooner sound like "Frank Sanatra" it gave me that message saying it can't reproduce Frank Sanatras voice. It said it replaced Sanatra. But it did end up sounding exactly like Frank S.

You don't have to copyright what a model is trained on. I play the keyboard, I train on Kate Bush, Radiohead, Led Zeppelin. I play guitar, I train on Jimi Hendrix, and Led Zeppelin. It doesn't mean anything so long as I don't recreate their music, and sell it. In fact to avoid a copyright you need to first know the copyright material. Don't confuse training with output. The output should avoid copyright material, but the training doesn't have to.

also the 80s was a decade of wonderful creativity and inovation.

An absolute fantastic research job here! This is a super meaningful analysis and needs to be shared widely

Actually sounded more like Paul than John, but either way sacrilegious!

As a Reaper user, I recently installed a Python script called 'Composer Assistant' on my Reaper DAW. This script harnesses AI neural-based technology to generate various versions of the original melody in MIDI format. Interestingly, a similar process seems to be unfolding here, albeit with audio. Given that every song can be deconstructed into its individual instrument parts, it becomes quite straightforward to craft alternative versions. Simply reassemble these parts, and presto! However, it's crucial to recognise that this practice is illegal. Despite this, the company behind such tools may attempt to evade legal repercussions, perhaps due to the somewhat lawless nature of the current technological landscape.

Dude the All I Want for Christmas is crazy :D But I thought the person who generates the track prompts the lyrics?

Pandora's box has been opened, you're fired. If Udio or Suno are denied this, some other company will take care of it. This is not just about music, the world is going to change. Adapt or die

Of course it’s trained on copyrighted music (without permission)

I love the "bold, inventive, originale and more creative" approche as a strategy to compete with AI generative music models in the future. I can't wait to learn how to apply that approche to production music. Do you think that could work for music licensing?

If you can get a non-copyright tune out of the AI then it follows that you could get a copyrighted tune out of it too if all of the training data is treated in the same way. People always seem to find ingenious ways around the attempts to prevent this happening by crafting 'special' prompts. It doesn't help that the creators of the AI don't know exactly how it works.

My idea is they are using something like moises to separate the stems and its training it on the stems.. And it using some sort of AI that judges the combined stems to determine if it sounds good, the copyrighted music is being used to train thee AI that judges the quality of the song, and its using the users likes and choices in the music making process to rate the choices, which is key to training the aesthetics of the music..

I m from France. Some Udio french songs voice really looking like an artist called Barbara

by the way even if you think the record industry will slap them down, much of it will make its way to open source which makes no money, and can't be removed cause of things like torrent and the fact the music industry as a whole is not significant to fight a world that may fight back '

Thanks for sharing ! I make music now for myself and not for sync anymore. I'm depressed to see this things. Udio en the others are gonna make a lot of money.

Our colleague has uploaded several thousand German big band records from the period between 1920 and 1940. When we generate songs like this, we listen to the old records again. But even when it comes to pop music, nobody dares to pick up a song. You always get the feeling that the song already exists in a similar way.

i saw an ads using one of great artist voice from asia to promote a product in RU-vid short add. It included a video when she was being interviewed in a show and i saw her mouth saying different things then what the ads does. but it's just an ads

Very interesting video. Much enjoyed your opening to this debate. As a time served (40+ years) producer of music here are my thoughts - 1) Licensing will come when the matter is tested, either in court or as a result of a threat of court. 2) If Udio (or any of the others) have any brain cells they will either - already have the licensing sorted, or are anticipating the need for licensing. 3) Fair use doesn't apply, for the reasons you outlined. 4) Udio in particular IS capable of original thought. It CAN take a relatively complex prompt and produce things that have not been heard before (at least by me). The four most recent tracks on my channel illustrate this. 5) Musicians who currently make a living from selling their music as a commodity, for sync purposes or for corporate use, are probably doomed. If you are a creative original, and you don't rely on your music to pay the bills you'll probably be OK. Indeed, you might find your work even more sought after. 6) This one's just an opinion, but creative people must embrace the AI and use it in a way that helps them. If they don't they risk being left behind. I'm thinking of AI applications within the studio/DAW environment. Great video, and very interesting times.

Dropped a bomb here, man - great work 👏👏 Music lawyers are gonna be busy

That clause doesn’t say Udio or suno don’t store music data in their data models. That’s a policy from OpenAI saying that OpenAI doesn’t copy or store data sent to it by customers for processing. That doesn’t mean the customers don’t store the data. Of course they do otherwise it wouldn’t work.

'Why is this happening for me' noted and locked in.. Thanks.

But to be fair, if it is a certain individual who is purposefully exploiting the website to generate sound-alikes despite the website explicitly attempts to make it impossible, shouldn't it be the individual to blame?

🤔 💭 uncanny. Udio got a problem. But when it comes to laws and changing laws big business can buy off court officials to get laws passed in their favor

5:43 whilst the vocals do not sound like Mariah it is sang in the same way we are all familiar with, so you can question how it was trained by all means but as per another comment i have made if the record studios have made it available to listen to for free on youtube, then its available to learn from, its note worthy that AI using various infrastructures to learn from to identify patterns and saves a fingerprint of the pattern, that being not the audio its self, but an encoded version of what makes it that style of audio, in a similar way to how we can identify a dog in a picture and its breed, by studying and learning from various pictures of dog breeds the patterns that each dog has ie German Shephard has the typical mole on their face, dark pronounced nose and the golden colours and black tones, but ai uses data augmentation so it will change the look of the image using filters to identify patterns such as RGBMAX as an example, which makes the image look totally alien and horrendous to look at but its establishing patterns and its the geographical patterns it stores, in this case its the wave hight, pitch, and speed that would be considered as patterns. Whilst people are quick to shout "COPYRIGHT" I ask this.... How is that any different to a cover version of a song ie: "The Police - Every Breath You Take" ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-OMOGaugKpzs.html vs "IMY2 - Every Breath You take" ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-N8HXe3INqKE.html Same title, same lyrics, similar instrumental, acapella is different but sang in the same way If we are going to sue big tech companies for this then we have to go after every single individual doing covers in bars, pubs, restaurants, academically, reality TV shows like Americans Got Tallent, and of cause upcoming artists using covers of existing songs

not sure but i tried it and it detects the lyrics and it wont allow you to create

I was inspired to go onto Udio and create a pop song about wondering if Udio is using copyrighted music. For its title, I gave it this youtube link, haha!

SOLID GOLD!! somebody's got some 'splainin' to do

Ya puedes por favor contar los promt para que suene como nos gustaria

are you sure you did not write the lyrics in when you created the loop

Excellent

well done !

They are accentuating the concept of a modicum of creativity.

Do the search you will find the videos.. I havent heard these precise tunes, but I know its got to be inpainting.. I think what is being done there is they are inpainting a segment then they give the AI enough of a context window (which can be controlled) to give it an idea of what to fill the rest of the song in with.. I just didn't know why until I found your videop.. Will I use it? No, its more fun making new songs with the artists by accident..

Wow!!!

Here is a related storyline that I (ironically) wrote with the help of AI. Once upon a time, in a world where AI had seeped into every facet of daily life, a group of artists known as the "Ephemerals" emerged. Tired of seeing their creative expressions reduced to mere training data for machines, the Ephemerals sought to create art that was impossible to digitize or replicate. They staged performances in secret, urban locations, leaving no trace of their work except for the memories they etched into the minds of those lucky enough to witness their art. From flash mobs that dissipated as quickly as they formed to intricate installations made of ice that melted in the sun, the Ephemerals' art defied categorization and eluded the grasp of AI. As their fame grew, the Ephemerals became symbols of defiance, their art a testament to the power of human creativity in the face of technological homogenization.

Been using Udio for two weeks, had it happen once, so personally I am concerned, obviously when a user who is musically trained and music savvy, they would obviously steer clear, however, I see issues arising, I don't think they did so intentionally, but yeah it is an issue. I am sure it will be phased out.

That doesn’t mean that that came off Udio, you could make a song on audio with a voice and then use voice replacement software to replace the voice with John Lennon‘s voice that was trained on something like replay app

AI really brings out the fuddy-duddy in people.

I used lyrics I wrote for a song that was never finished in my home studio. Udio frighteningly got it close to the original idea…

There's a company that makes realistic sounding woodwind and brass samples for composers. They use a hybrid technique where they merge modeling with recorded samples. Probably that same analogy is what is happening on Udio. I've had situations where I don't like what Udio is making so I keep regenerating the same section, and you can tell it is making original phrasing and such based on what you enter. Like I'm entering ChatGPT written lyrics, so Udio has nothing to reference in those cases, yet it can still create an adequate melody based on the syllables in each line. So it is producing original content in that respect. But it is suspicious that it will recreate well known songs too. They need some type of filter like RU-vid or SoundCloud has.

I made one that sounds like David Coverdale from Whitesnake. Yeah they're just letting the model keep these original voice timbres. They should do an update where they morph these famous voices with non-famous voices. Maybe do something like eleven labs that pays you to use your voice.

hello, me again lol, i think you could do shorts of the most important evidences on this video, shorts are highly viral ;)

that was Mariah Cariah on 'All I want for christmas', that was definitely Paul Mcartney on the second 'Beatle' tune, there is no getting around it.

Well, they've said that you aren't allowed to use any of the use commercially, so that makes sense

Of course the model contains copyrighted music. Now, the Copyright infringement happens when you put the same lyrics and trick the prompt to get the artist anyway. It's up to the user, like with every tool, you are responsible of how you use it. Also, with the inpaint feature, you can torture a song until it makes the exact chord progression you want.

Neural networks does not work in the way that they store actual music and instead works more like how the human brains stores memories. What I suspect is going on after experimenting with training models myself is that it has been overtrained on too few samples/songs. This would not be the issue if it was trained right on larger variety of songs within each genre and more variety would also mean more creativity.

My concern here is that there is no indication of whether the LYRICS were fed to Udio or not. I don't know the lyrics to all these tunes, but if they do match, then maybe the lyrics were fed. And if so, that's not exactly a fair examination, is it?

You are totally correct, AI can never be really creative (although I'm sure it can appear to become rather close). Much of the music that inhabits the Spotify etc "air waves" is essentially imitative and AI can replace so much of that material. The same can be said for music for film/TV/commercials/business etc etc. In the short term, this could/will put many performers/producers/bands etc out of work. But people ready to show originality, to be creative, will progressively bloom. And live music - any genre - will never die. It's an "industrial revolution", and can't be stopped. A simple parallel is the impact the widespread introduction of electricity had on society. I think the first part of this posting is pertinent, but it's issues are really transitory, they will be resolved, disappearing into the past. On the other hand, the latter part of your presentation is something everyone should hear. In essence, your principle of positivity, of taking circumstances where each would be seen as 'a negative' and turning the circumstance on its head to see a positive side, and so use it to move forward.. i like your thinking.

Wow. That is nuts.

setting aside right or wrong, the reason it's able to reproduce song very similar to the ones you showcased here, is most likely because these are all songs that have already been covered and re-recorded ad nauseum. They clearly didn't filter covers and alternate versions from their data set, and so the AI ends up spending more "time and attention" on these very similar recordings, and learns the exact patterns in those songs more accurately. The other reason is they've had to weigh some material more than other - meaning, a lot of published music isn't actually "good", and so, after training on "everything", they've almost definitely finetuned the AI to more close match the patterns in popular music and hit songs. In order to produce something most people will like, they have to do that, otherwise 90% of it's output wouldn't be something people like. we've seen similar issues with other forms of AI - for example, when ChatGPT was able to reproduce portions of New York Times articles, because those were weighted higher. just as OpenAI did, Udio will likely be forced to comment on this at some point - and they will probably say what OpenAI said, that this is a "bug" in the model. unfortunately, this is not a "bug" they can fix - you can't really "remove" things from a trained model, because that would "damage" or "untrain" other aspects of the model and lower the quality, we've seen this as well. You can't go back and adjust your training set either, unless you retrain the entire model from scratch, which would would likely cost tens of millions and take months. as long as this industry is unregulated, they will keep doing what they want - if it were regulated, they would likely be left with two options: shut down or lie. cleanly curating a fully-licensed useful data set is nigh on impossible in terms of either cost, effort, or data availability. whatever you think about all of this, if it were more regulated, no (legal, compliant) version of either Udio or ChatGPT would even exist.

If you put the exact lyrics of a song that already exist into Studio then it's pretty likely eventually if you cherry pick your results you'll get the song the lyrics from. But if you put your own lyrics in Studio Will do what your brain does, it takes all the music you've ever heard and applies it to the idea you have

There is no doubt that they have trained the AI with all music on the internet. I have already reproduced the voices of artist of many genres inadvertently just by using common prompts for the artist's actual genre. Subbers who have more gen opportunities will most likely find even more (im free) and the simplest way to find this is by doing this with the actual lyrics to a song. Just as this video showed. They will in the future be doing this to artist who provide their songs in their site for "AI training" that is all you need to know to understand any confusion with what is happening.

Okay, hi! I'm Lynn Cole. New subscriber. I'm an experimental artist. Suno got me hooked on making music, so I bought a keyboard and a daw. Now I'm an experimental musician and producer too. I have two kinds of music I make these days, 1. Sweary country music about drugs and toxic love affairs. 2. Trying and failing to be Frank Zappa. If you saw my sound cloud, you would understand. My lyric videos are okay, but SC's really the place to track my journey as a songwriter. I have some thoughts here. First and foremost, you have the right approach and attitude. You do. That's because you're an artist, and you get it. Next, Red Lobster used Suno, during a period of time where the model was REALLY fried, and you can hear it in the tracks. If you pull it up with a spectral, you'll see that the mid pass is all messed up. That's suno v3p1. They're on either p2 or p3 right now. V4 is coming, or should be. If it were me, and I was suno, it would be coming a month ago. That mixtape was awful though. It's more than cringe worthy, and everyone who's looked at it and said, "wow, I hate it." You're right. It's a zero effort campaign, and it shows. You noticed it, their customers noticed it, everybody noticed it. Does that mean an end to sync fees? Dude, I dunno, how scared are you that the people who buy your work have no taste? I feel like, if you're ever worried about fully ai generated music. Just go to suno or udio and tell it to generate an entire song with no human intervention, editing, or songwriting, and see what comes out. I got news for you. If you're impressed by what comes out lyrically, you have absolutely no business calling yourself a songwriter. None. Just go home, take your neon light cliches, and call it a day. You did good, you'll be replaced. If you're not, congratulations! You're probably at least adequate, and don't suck at your job! Achievement reached. As someone who tells herself she's making serious art, I want the legal consensus in the US to be reached on this. Right now, I'm only famous in China, and that sphere of markets. I would like to be taken more seriously by the music industry, and the only way to achieve that is by legal certainty. Regardless of the outcome in the courts, I support it, and look forward to what happens. No matter what, it's going to be good for everybody. Also, please post hashes for those tracks so I can take a look at them. Would love to share this with Udio support.

Important questions asked. Two thoughts: 1) I don’t think the sole creation of a piece of music is a problem, even if it’s an identical copy of Last Christmas. Nothing stops me from making a copy of a well-known song. What I do with it is another question. 2) To your point regarding human innovation as a way to compete: I’ve heard Udio spit out all kinds of weird stuff, sometimes barely music, that I’ve never heard before, including insane genre fusions and/or unidentifiable sounds and instruments. Wonder if that can be considered more innovative than many music producers out there including those who'd say their intent is to innovate.

But what people are saying is adobe is working on a tool that will be able to do what UDIO is doing with any music you give it..