Ed Newton Rex, the former VP Of Audio at Stability AI who resigned over the company's position on fair use, has launched Fairly Trained, a new initiative that will certify generative AI companies with “training data practices that respect creators’ rights". The scheme has the backing of a number of copyright owners, including music companies Universal Music and Concord.
There is an ever-growing dispute between the tech sector and the copyright industries, including the music industry, over the copyright obligations of AI companies. The music industry is adamant that if a company trains a generative AI model with existing music, they must get permission from the relevant copyright owners.
However, many tech companies argue that AI training is covered by copyright exceptions in at least some countries and, under US law, constitutes fair use, which means that no permission is required. That includes Stability AI, even though its music product, which Newton Rex oversaw, was training with licensed music.
"There is a divide emerging between two types of generative AI companies”, says a blog post launching Fairly Trained, “those who get the consent of training data providers, and those who don’t”. However, “right now it’s hard to tell which AI companies take which approach".Fairly Trained, it goes on, aims to change that by offering a number of certifications.
The first to be made available is called Licensed Model Certification, “which can be awarded to any generative AI model that doesn’t use any copyrighted work without a licence”. At launch, nine AI companies have been awarded the mark, covering image, music and singing voice generation. They are: Beatoven.AI, Boomy, BRIA AI, Endel, LifeScore, Rightsify, Somms.ai, Soundful and Tuney.
Welcoming all this, Universal Music's SVP Strategic Technology Christopher Horton says: “Universal Music Group has taken an industry-leading position in calling for responsible and ethical uses of AI technology that further enhance creativity, support artists and respect copyright. We welcome the launch of the Fairly Trained certification to help companies and creators identify responsible generative AI tools that were trained on lawfully and ethically obtained materials".
The Fairly Trained blog post also sets out some of the initiative's future objectives. While the initial Licensed Model certificate is relatively basic, other certificates will be developed to deal with some of the complexities. For example, "to address more nuanced questions about dataset acquisition, such as opt-in vs opt-out".
Within the music community, artists and songwriters agree with record labels and music publishers that AI companies should get permission to use existing music to train models. However, they also want their labels and publishers to secure consent from each music creator before allowing their recordings and songs to be used to train AI, and would likely argue that that commitment is also required to truly respect creators' rights.
It remains to be seen if that kind of commitment could also be confirmed by a certification programme.