Lyrics produced some of the first AI slop I noticed after ChatGPT was launched in late 2022, even if the large models hadn’t been trained on them specifically. Overnight there were a bunch of different advertising-laden sites that clearly scraped Genius or other lyric websites, and then had GPT generate commentaries on what the lyrics supposedly mean, so that these would get picked up by search engines.
The result was mostly comical, the commentaries for vacuous pop music all sounded more or less the same: “‘Shake Your Booty’ by KC and the Sunshine Band expresses the importance of letting one’s hair down and letting loose. The song communicates to listeners how liberating it is to gyrate one’s posterior and dance.” Definitely one of the first signs that this new tech was not going to be good for the web.
Member when music sites were suing YouTube for music videos, and now they are begging people to watch them there and YT view counts are a bragging topic?
Soon music industry will be begging OpenAI for exposure of their content, just like the media industry is begging Google for scraping.
Of course the models are not human, but if you consider this situation as if they are persons, then the question becomes: May a person read lyrics and tell it to someone when asked, and the court's ruling basically says no, this may not happen, which makes little sense.
I guess the main difference between the situation with language models and humans is one of scale.
I think the question should be viewed like this, if I as a corporation do the same thing but just with humans, would it be legal or not. Given a hypothetical of hiring a bunch of people, having them read a bunch of lyrics, and then having them answer questions about lyrics. If no law prohibits the hypothetical with people, then I don't see why it should be prohibited with language models, and if it is prohibited with people, then there should be no specific AI ruling needed.
All this being said, Europe is rapidly becoming even more irrelevant than it was, living of the largess of the US and China, it's like some uncontacted tribe ruling that satellites can't take areal photos of them. It's all good and well, just irrelevant. I guess Germany can always go the route of North Korea if they want.
With AI slop showing up everywhere, there’s a real danger that folks will just no longer be motivated to produce real original content.
With all major models not basically trained on nearly all available data, beyond the financial AI bubble about to burst there’s also a big content bubble that’s about exhausted as folks are just pumping out slop vs producing original creative human output. That may be the ultimate long term tragedy of the present AI hype cycle. Expect “made by a human” to soon be a tag associated with premium brands and customer experiences.
I am curious what happens if they call their bluff on this and cut off ChatGPT in Germany. Not that I think OpenAI is doing the right thing, just, I don’t think a country’s government can justify no commercial LLMs to its populace.
> Since the output would only be generated as a result of user inputs known as prompts, it was not the defendants, but the respective user who would be liable for it, OpenAI had argued.
Another glimpse into the "mind" of a tech corporation allowing itself full freedom to profit from the vast body of human work available online, while explicitly declining any societal responsibility at all. It's the user's fault, he wrote an illegal prompt! We're only providing the "technology"!
I m not sure about the problem here, lyrics are public you can search '$songname lyrics' and get the result in a website (or even at the search engine page). What's the issue with an LLM producing those lyrics if you ask?
While I partially understand (but not support) the hate against AI due to possible plagiarism and "low effort generation" of works, think about the whole process: If model providers will be liable for generating output, that resembles lyrics or very short texts that fall under copyright laws, they will just change their business model.
E.g. why offering lame chat agents as a service, when you can keep the value generation in-house. E.g. have a strategy board that identifies possible use cases for your model, then spin off a company that just does agentic coding, music generation. Just cut off the end users/public form the model access, and flood the market with AI generated apps/content/works yourself (or with selected partners). Then have a lawyer checking right before publishing.
So this court decision may turn everything worse? I don't know.
These people would stream German schlager to every screen and speaker in Europe and charge for it 100 EUR monthly per breathing person, if they could. They are violent.
However, the lyrics are shown because the user requested them, shouldn't be the user be liable instead? The same way social networks are not liable for content uploaded by users? I think here there is a somewhat double standard.
Of course, maybe OpenAI et al should have get a license before training on the lyrics or to avoid training on copyrighted content. But the first would be expensive and the latter would require them to develop actual intelligence.
Everyone knows that these LLMs were trained on copyrighted material, and as a next-token prediction model, LLMs are strongly inclined to reproduce text they were trained on.
post trained models strongly inclined to pass response similar to what got them high RL score, it's slightly wrong to keep thinking of LLMs as just next token predictions from dataset's probability distribution like it's some Markov Chain
Another instance of GEMA fighting an american company. Anyone who was on the german internet in the first half of the last decade remembers the "not available in your country" error messages on youtube because Google didn't make a deal with GEMA.
I don't think that we will end up here with such a scenario: lyrics are pervasive and probably also quoted in a lot of other publications. Furthermore, it's not just about lyrics but one can make a similar argument about any published literary work. GEMA is for music but for literary publications there is VG Wort who in fact already have an AI license.
I rather think that OpenAI will license the works from GEMA instead. Ultimately this will be beneficial for the likes of OpenAI because it can serve as a means to keep out the small players. I'm sure that GEMA won't talk to the smaller startups in the field about licensing.
Is this good for the average musician/author? these organizations will probably distribute most of the money to the most popular ones, even though AI models benefit from quantity of content instead of popularity.
The obsession with protecting access to lyrics is one of the strangest long-running legal battles to me. I will skip tracks on Spotify sometimes specifically because there are no lyrics available. Easy access to lyrics is practically an advertisement for the music. Why do record companies not want lyrics freely available? In most cases, it means they aren't available at all. How is that a good business decision?
Thoughts by someone who doesn’t make a living by songs?
I’m guessing you’d want to restrict lyrics to encourage more plays of the song by people who are motivated to understand them. Along with the artist’s appreciation of that experience of extracting what you’re fascinated by. Burdensome processes generate love and connection to things.
Not everything is a functional commodity to be used and discarded at whim.
Can't they just ask for copies of the lyrics they are not allowed to use and s/lyrics//g the training set? I imagine the volume of text that will be removed would be relatively miniscule.
I think in the end they will just pay off copyright holders. The German GEMA is mostly interested in rent-seeking through whatever means available, it's basically the whole point of the organization.
They'll easily be paid off once all legal avenues are exhausted for OpenAI. Though they'll of course keep fighting in court in the hopes of some more favorable negotiating position.
Copyright law continues to stifle innovation. The DMCA needs to be abolished, and we need to entirely rethink our modern economic system with respect to creative industries (including software development). The cat is not going back in the bag.
We are sitting on the precipice of the greatest technological advancement in history, and rent-seeking industry titans have convinced us that we must stop this unstoppable technological advancement in order to protect the livelihood of artists who already receive cents on the dollar for their efforts.
Doesn't that sound familiar? This is what they have done time and again, and each time they have lost, leading to a huge loss of potential revenue for creatives as people make use of technological breakthroughs.
Then when companies like Netflix finally get everyone on board with streaming and paying for content with modern conveniences, industry titans step back in to demand larger slices of the pie, until the entire system is ruined and people return to piracy and consumption of older media. Don't even get me started on Spotify.
The technological benefit of modern machine learning models is just too large to ignore. These are becoming important tools, which put power back in the hands of the people, of the consumer. A lot of the grassroots anti-AI movements we see in the creative space can be traced back to corporate propaganda or financial backing. A lot of these people really think they're doing what's best for artists. But I just see Blockbuster all over again. We should make an effort not to be on the wrong side of history.
30 comments
[ 4.1 ms ] story [ 58.2 ms ] threadThe result was mostly comical, the commentaries for vacuous pop music all sounded more or less the same: “‘Shake Your Booty’ by KC and the Sunshine Band expresses the importance of letting one’s hair down and letting loose. The song communicates to listeners how liberating it is to gyrate one’s posterior and dance.” Definitely one of the first signs that this new tech was not going to be good for the web.
However of course OpenAI will ignore this and at worst nothing will change and at best they get a slap on the wrist and a fine and continue scraping.
You can’t take that stuff out of the models at this point anyway.
Soon music industry will be begging OpenAI for exposure of their content, just like the media industry is begging Google for scraping.
I guess the main difference between the situation with language models and humans is one of scale.
I think the question should be viewed like this, if I as a corporation do the same thing but just with humans, would it be legal or not. Given a hypothetical of hiring a bunch of people, having them read a bunch of lyrics, and then having them answer questions about lyrics. If no law prohibits the hypothetical with people, then I don't see why it should be prohibited with language models, and if it is prohibited with people, then there should be no specific AI ruling needed.
All this being said, Europe is rapidly becoming even more irrelevant than it was, living of the largess of the US and China, it's like some uncontacted tribe ruling that satellites can't take areal photos of them. It's all good and well, just irrelevant. I guess Germany can always go the route of North Korea if they want.
With all major models not basically trained on nearly all available data, beyond the financial AI bubble about to burst there’s also a big content bubble that’s about exhausted as folks are just pumping out slop vs producing original creative human output. That may be the ultimate long term tragedy of the present AI hype cycle. Expect “made by a human” to soon be a tag associated with premium brands and customer experiences.
> Since the output would only be generated as a result of user inputs known as prompts, it was not the defendants, but the respective user who would be liable for it, OpenAI had argued.
Another glimpse into the "mind" of a tech corporation allowing itself full freedom to profit from the vast body of human work available online, while explicitly declining any societal responsibility at all. It's the user's fault, he wrote an illegal prompt! We're only providing the "technology"!
E.g. why offering lame chat agents as a service, when you can keep the value generation in-house. E.g. have a strategy board that identifies possible use cases for your model, then spin off a company that just does agentic coding, music generation. Just cut off the end users/public form the model access, and flood the market with AI generated apps/content/works yourself (or with selected partners). Then have a lawyer checking right before publishing.
So this court decision may turn everything worse? I don't know.
Of course, maybe OpenAI et al should have get a license before training on the lyrics or to avoid training on copyrighted content. But the first would be expensive and the latter would require them to develop actual intelligence.
Everyone knows that these LLMs were trained on copyrighted material, and as a next-token prediction model, LLMs are strongly inclined to reproduce text they were trained on.
I don't think that we will end up here with such a scenario: lyrics are pervasive and probably also quoted in a lot of other publications. Furthermore, it's not just about lyrics but one can make a similar argument about any published literary work. GEMA is for music but for literary publications there is VG Wort who in fact already have an AI license.
I rather think that OpenAI will license the works from GEMA instead. Ultimately this will be beneficial for the likes of OpenAI because it can serve as a means to keep out the small players. I'm sure that GEMA won't talk to the smaller startups in the field about licensing.
Is this good for the average musician/author? these organizations will probably distribute most of the money to the most popular ones, even though AI models benefit from quantity of content instead of popularity.
https://www.vgwort.de/veroeffentlichungen/aenderung-der-wahr...
I’m guessing you’d want to restrict lyrics to encourage more plays of the song by people who are motivated to understand them. Along with the artist’s appreciation of that experience of extracting what you’re fascinated by. Burdensome processes generate love and connection to things.
Not everything is a functional commodity to be used and discarded at whim.
They'll easily be paid off once all legal avenues are exhausted for OpenAI. Though they'll of course keep fighting in court in the hopes of some more favorable negotiating position.
We are sitting on the precipice of the greatest technological advancement in history, and rent-seeking industry titans have convinced us that we must stop this unstoppable technological advancement in order to protect the livelihood of artists who already receive cents on the dollar for their efforts.
Doesn't that sound familiar? This is what they have done time and again, and each time they have lost, leading to a huge loss of potential revenue for creatives as people make use of technological breakthroughs.
Then when companies like Netflix finally get everyone on board with streaming and paying for content with modern conveniences, industry titans step back in to demand larger slices of the pie, until the entire system is ruined and people return to piracy and consumption of older media. Don't even get me started on Spotify.
The technological benefit of modern machine learning models is just too large to ignore. These are becoming important tools, which put power back in the hands of the people, of the consumer. A lot of the grassroots anti-AI movements we see in the creative space can be traced back to corporate propaganda or financial backing. A lot of these people really think they're doing what's best for artists. But I just see Blockbuster all over again. We should make an effort not to be on the wrong side of history.