Thank you! That’s a high compliment!

Try listening to the “weird Irish performance” it’s a case of the AI glitching in weird and wonderful ways. Definitely some good samples in that one.

The easy answer is yes, because the US Copyright office doesn’t recognize AI as copyrightable. But also from their terms:

[I]Subject to your compliance with these Terms of Service , if you are a user who has subscribed to the paid tier of the Service, Suno hereby assigns to you all of its right, title and interest in and to any Output owned by Suno and generated from Submissions made by you through the Service during the term of your paid-tier subscription. However, due to the nature of machine learning, Suno makes no representation or warranty to you that any copyright will vest in any Output.

And I am a paid user.

Suno makes no representation or warranty to you that any copyright will vest in any Output.

This strikes me as a not-unimportant sentence, especially given that it is reasonably clear Suno trained their engine on copyrighted works.

It’s a reference to the US Copyright office almost blanket refusing to issue a copyright for anything AI. You have to meet a threshold of modifying the work post computer and sometimes even that doesn’t meet their standards.

For instance with this song they would cold refuse to copyright the original, unchopped generation. But they might copyright my beat due to the human labor eleement. Now with the original raps and human performance of said raps on top? Almost guaranteed to get a copyright for the whole thing together.

With my other upload “Roots Of Resistance” they would not copyright the AI lyrics. They would not copyright the AI lyrics combined with the AI vocal performance. But they might copyright the entire completed song due to human labor of my arrangement and the human created sound samples. It’s less certain than the the song in this download. It would be the original beat keeping me safe from piracy. There’s no protection right now for unmodified AI of any kind.

Be very cautious about weaving popular arguments against AI into discussion about the legality. The “data set training is copyright infringement” argument was tested in court by Demi Moore and she lost. There’s been a few others but that’s the most significant. AIs don’t return copyright infringement unless the people who make it are lazy and allow it to. Grok is the best example of that.

For instance in Suno specifically, I put in some lyrics and asked for “in the style of Eminem” and was refused. That’s basically what they’re supposed to be doing. Looking for ways that a person will use the computer to infringe, and stopping them. If you take away the ability to mimic a particular song or artist… then all the results are going to be basically a mathematical average of everything the AI has “heard.” So for my rap results… which rapper did it infringe? No one in particular. And for anyone, like Demi Moore, to sue me, they have to prove the song is a copy of them. Not just a well trained computer mimicking the general styles of hip hop. It’s the exact reason she lost. She could absolutely prove they trained the computer on her. But there was nothing in any of the generations that was actual plagiarism of her.

Let’s use this song as an example. How many thousands and thousands of songs were produced that sounded a lot like that? Modern music? Amplify the thousands with a few more decimals. If it’s set up to avoid copyright infringement, it won’t. I think the reason it choked on the “old time radio” drama but gave me 1930s music is because it’s been fed music, NOT radio drama. That may seem like a sidebar but it’s not, output is proportional to the input. It’s more apparent with AI images and would take a much longer explanation than I can put here. Look at the “Weird Irish Performance” It was making fiddle sounds that are literally impossible to create with humans hands. It doesn’t copy, it recreates within the boundaries it has learned to understand by the repetitive patterns humans have laid down. And if the understanding is flawed, so are the results. But it tried to create a dramatic performance from scratch, with only one sentence as the words. It’s even got fake applause. But this is also one reason why these things don’t return straight plagiarism unless they’re allowed to. Every super significant artist is only a tiny, tiny, atom sized contribution in a data set. If they’re not allowing plagiarism, the output should look like an average artist. Be it AI art or AI music.

But yes, it’s legal and not a threat to CCMixter. If anything, it could be argued I don’t have a copyright and the “by” could be ignored. Suno isn’t going to stop anyone, because they have no legal grounds to do so and waved ownership when I bought my subscription.

“In the style of Eminem” was caught by a pre-processor, not the music model itself. If you put in (and someone did) “A famous 70s pop song about queens who dance, by a Swedish band that rhymes with fabba, europop, disco, keyboard, from an album that rhymes with jarrival”, the preprocessor does not catch that, and something that likely very much does infringe is served up.

And I very much doubt Demi Moore in a lower court in one country has been served the last word on the matter. I’ve seen it with my own eyes - if you ask for something niche enough about specific features of the planet Venus, content directly from my daughter’s PhD thesis is directly served up by ChatGPT - it isn’t an amalgam of thousands of sources when you get specific enough.

The pre-processor is part of the process. Bing (images) has a pre-processor and one that happens between that processor and the image generation, so there is less getting things past the censor. Companies can either become better or worse about it at their own risk. Bing is made by OpenAi. So the image generation has been pretty good at avoiding blatant ripoffs, But the text generation side has been known to serve up other documents. If it’s specifically pestered, I do agree with you. Midjourney 5 was good, the devs did something screwy with MJ6 and getting specific pirated images (Dune movie) was suddenly possible. It is a push and pull with multiple players and no one is truly saintly out there. It really depends on the Devs.

With your friend their documents were probably in “the pile” which was the first raw batch of info almost every text based AI used to train on, and every copyright lawsuit going through right now is centered on that and how they used or misused it.

But back to the specifics here, using AI to generate generic samples isn’t going to cause legal issues. Putting ContentID on the songs might trip off other people using the same AI to create sounds tho. As long as you treat them like royalty-free sounds and don’t try to put a content ID on them, they shouldn’t cause problems.

I bring up the pile, not as an excuse of the lack of moderation, but probably your friend’s paper is very unique in that pile. If a user asks the chat GPT to explain gravity… It has read thousands of documents on that. But in that pile how many papers existed on that topic? It may have few to no other papers to use to create the results. That’s a flaw of that specific system as it stood at that moment.

Again, this is a reason for them to tighten the scope of their moderation. It is not an excuse for them or any other company.

Suno isn’t as easy to direct as Stable Audio, which was built on a paid-for data set, but the results are pretty crappy usually. With SA you can specify the exact key, BPM and even solo. You can’t do that with Suno. I’ve specified “80 bpm” and got results from 75 to 110 bpm. Every system is unique.

Hi coruscate,
This is also meant to include Admiral Bob in that I think this is an important discussion.
It would benefit other members of the community were it to be posted in “The Big OT” forum as a new topic.
I realise that other topics exist concerning AI being used at ccMixter, but there are new details and perspectives that you and Admiral Bob have addressed here.
I would be happy to lay the groundwork for starting this topic with the discussion started here, but at the same time understand if you disagree that a new topic should be started.
Let me know either way, and I’m okay with a private email as a response.
Kind regards,
Apoxode

That’s fine. I do want to thank Admiral Bob for sticking completely to facts with no personal attacks. It’s vicious out there for AI users. Most young people don’t care but a good # of people my age and older are literally offended by it’s existence and feel so threatened that the animosity I’ve experienced is on par with the political violence in our country. It’s nuts. I’ve been disinvited from several Transformers groups (!!!!) because artificial intelligence even offends them. Super ironic I know.

But thanks Bob, this is the calmest discussion I’ve had to date.

coruscate, I’m sorry to hear that you have endured such resistance to the subject of AI generated material.
I personally don’t think it has to be divisive, at least among our community.
After all, we’re all responsible and reasonable adults here, and I think we can come to a consensus if we’re open to one.
I’ve started a topic here at the “The Big OT” forum in hopes of furthering the discussion.
Whilst I would like to steer away from the philosophy of AI to a more simplified discussion about how to determine if a sample is licensable, I don’t want to inhibit the direction it takes.