- A recent controversy at Midjourney saw 16,000 artists’ styles being used without consent to train image generation models.
- Copyrighted data is widely used to train language and image AI models.
- What might this mean for AI in music? What would happen if text-to-image audios began spitting out well-known riffs and melodies?
Over in the visual arts, Midjourney, who developed a text-to-image AI capable of generating incredibly detailed art, was blasted after a leaked spreadsheet with over 16,000 artist names was leaked online.
Meanwhile, OpenAI, the company behind ChatGPT, said that training these kinds of AI models without using copyright data was “impossible.”
That’s a pretty seismic statement that provoked some pretty intense debate on social media, with onlookers saying this shows how OpenAI’s business model relies on, well, theft.
After all, if you or I produced a product featuring Mickey Mouse or Woody from Toy Story, there would be hell to pay.
Now, consider these AI companies are worth billions – it’s easy to see why people are mad about them using copyright data so freely, without coughing up a dime for creators thus far.
So, what does this mean for music AI?
It’s fair to say that text-to-audio AIs are more complex and have come on slower than image or language-based models, but with LimeWire, Google, and Suno releasing competent tools recently, it’s only a matter of time before we can request detailed pieces of music with short text prompts.
The music industry, renowned for its vigorous defense of copyright, will then enter uncharted waters where AI-generated music could inadvertently infringe upon existing copyrights, leading to complex legal battles reminiscent of those faced by Midjourney, OpenAI, Stability AI, and Anthropic, to name but a few.
Ownership of AI-generated music is another area of ambiguity. It’s unclear whether the person inputting prompts into an AI tool or the creator of the tool itself should be regarded as the author. This ambiguity mirrors the confusion in the visual arts sector, as highlighted by the Midjourney case.
Furthermore, proving copyright infringement in AI-generated music is challenging. AI tools are designed to emulate the general sound and feel of music, making it difficult to pinpoint specific instances of copying.
But what would happen if one of these tools were to spit out a very well-known riff? Or worse, even, lyrics.
Just imagine the lawsuits if these were to create something as recognizable as ‘Smoke on the Water’ by Deep Purple, ‘Dancing in the Moonlight’ by Toploader, ‘Shiver’ by Ed Sheeran, etc, etc.
AI is already reshaping music
The evolving landscape of AI in the music industry is shifting how music is created, distributed, and protected under copyright laws.
For instance, an AI-generated verse mimicking Kanye West’s style raised questions about the limits of AI in replicating artist styles and the potential infringement on artists’ rights.
Perhaps there’s no more controversial case than Ghostwriter’s ‘Heart on My Sleeve,’ which featured Drake and The Weeknd – though it didn’t really – it featured an AI mash-up with them named on the track.
Unsurprisingly, Ghostwriter has struggled to monetize its work.
Looking ahead, the music industry faces difficulties in defining and protecting AI-generated content. The US Copyright Office has begun to address these issues, focusing on the role of human authorship in AI-generated works.
The new guidelines suggest that greater human involvement in creating AI-generated content might increase the likelihood of it being eligible for copyright protection.
However, the onus is on the creator to demonstrate the extent of their creative contribution to the AI-generated work.
Looking to 2024, developers of text-to-music AIs will need to keep a close eye on this year’s lawsuits, as if their models start churning out copyright riffs and melodies, there will be hell to pay.