When Meta recently unveiled its MusicGen music AI model, it explained that it had been trained on more than 20,000 hours of licensed music, including a catalogue from stock library ShutterStock and its Pond5 subsidiary.
Now ShutterStock has expanded its partnership with another leading light of the generative-AI sector: OpenAI, the company behind the ChatGPT and DALL-E services, as well as past music-focused research projects Jukebox and MuseNet.
Those latter two aren’t commercial services, but they sprang to mind when we read ShutterStock’s announcement of a new six-year partnership with OpenAI.
“As part of this expanded collaboration, OpenAI has secured a license for access to additional Shutterstock training data including Shutterstock’s image, video and music libraries and associated metadata,” explained the company [our emphasis].
Some thoughts. First, is this a hint that OpenAI has specific plans for AI music, or just about wrapping music into an expanded licensing deal for future possibilities?
In a recent appearance in the US Senate, during a grilling about training AIs on copyrighted music, OpenAI’s CEO Sam Altman was keen to play down its past music projects. He called Jukebox “a research release… not something which gets much attention or usage”.
Yet he also confirmed that OpenAI was engaging in dialogue with musicians as well as visual artists to shape its policies and approach. If any new music projects are brewing, being able to use ShutterStock’s production-music catalogue for training would be handy – just as it was for Meta.
Second, what does it mean that production music appears to be the go-to source of training material for these companies?
Music labels have been very clear in their belief that if AI models are to be trained on commercial music, they must be properly licensed. Yet stock-music firms are seemingly quicker off the mark to actually strike those deals.
That may be inevitable, given the relative simplicity of the rights around those companies’ music compared to the landscape of commercial recorded music and publishing. Still, it raises the question of a potential missed opportunity for both sides.
“The thing with AI is that the output is only as good as the input. You need high-quality data sets to train your models on, and this is where we come to a fork,” the CEO of AI music startup Endel told us in February. “Most of the AI music models were trained on just stock music, or stems that were created by a bunch of session musicians.”
In order for AI music to become as good as the actual music that we all love, it needs to be trained on actual [commercial] music,” he continued. “Otherwise your output will still sound like stock music.”
Companies in the production-music space, and the musicians who work with them, will understandably bridle at this kind of talk. But if stock-music companies are currently in pole position for partnerships to train musical AIs, it’s worth thinking about what that means.
(And, without trying to start a conspiracy theory: if this really does place a ceiling on the potential quality of AI-generated music, whether that’s something rightsholders are entirely unhappy about…)
Tools: platforms to help you reach new audiences
Tools :: Wyng
Through Music Ally’s internal marketing campaign tracking, we’ve recently discovered an interesting website by the…