(Bloomberg) —
A new initiative will evaluate and certify artificial intelligence products as copyright-compliant, offering a stamp of approval to AI companies that submit details of their models for independent review.
The certification is conducted by a nonprofit called Fairly Trained, founded by Ed Newton-Rex, who resigned in November as the vice president of audio for Stability AI, citing concerns over AI “exploiting creators.”
The Fairly Trained label will be given to AI companies that have obtained consent for the data they use to train AI systems, similar to how a tomato farm is certified as organic or a cocoa product is deemed fair trade. Fairly Trained has nine companies that have been certified so far, all small startups.
Getting permission to use copyrighted training data isn’t the norm. Generative AI leaders like OpenAI have argued that using data without a license falls under a “fair use” exception to US copyright law. Many rights holders disagree, which has led to several lawsuits, including a high-profile case in which New York Times Co. is suing OpenAI.
“All the training data has been stolen,” Salesforce Inc. Chief Executive Officer Marc Benioff said in an interview Tuesday at the Bloomberg House in Davos, Switzerland. Benioff also owns Time magazine, which is negotiating with OpenAI over a license agreement. OpenAI has struck agreements with the Associated Press and Axel Springer SE.
Newton-Rex quit Stability because he disagreed with the AI startup’s position on fair use — especially because he says that generative AI output competes commercially with the copyrighted works used for training.
Since resigning, Newton-Rex said he’s spoken with companies building models trained only on licensed or public-domain material, and he’s heard from businesses and consumers who would prefer to use models like that. He’s hoping that a Fairly Trained stamp can give people the ability to select models that align with their values.
“As long as generative AI companies are saying, ‘Yes, this is fair use, we can scrape whatever we want,’ then I think it will be a battle between the two sides,” Newton-Rex said. But plenty of companies aren’t saying that, and he said the issue doesn’t have to be so adversarial. He hopes to show that “there’s a mutually beneficial setup here that works for everyone.”
Here’s the list of Fairly Trained-certified startups:
- Beatoven.AI
- Boomy
- BRIA AI
- Endel
- LifeScore
- Rightsify
- Somms.ai
- Soundful
- Tuney
Eight are related to sound generation, and one creates images. To get certified, companies provide detailed information about their training data sources and make commitments around using licensed material. They pay a small submission fee and then, if certified, as much as $6,000 a year depending on their revenue, Newton-Rex said.
Advisers to the nonprofit include Maria Pallante, the CEO of the Association of American Publishers, a book publishing trade association; Elizabeth Moody, an attorney; Tom Gruber, a co-founder of the team that created Siri; and Max Richter, a composer and pianist (who OpenAI CEO Sam Altman recently mentioned as a composer he listens to often).
To contact the author of this story:
Ellen Huet in San Francisco at ehuet4@bloomberg.net
© 2024 Bloomberg L.P.
