Training models responsibly: AI and creator rights

The training of AI systems has sparked several ethical controversies, particularly concerning copyright, compensation, and fair use. AI models are often trained on vast datasets that include copyrighted material. There is controversy over whether using copyrighted content for training without explicit permission constitutes copyright infringement. Creators argue that their work is used without authorisation, potentially violating their rights.

AI companies and developers can monetise their models by offering AI services or selling generated content, potentially profiting from the uncompensated use of copyrighted material. The benefits of AI systems often do not extend to the original creators whose works were used in training, leading to calls for more equitable distribution of the economic gains derived from these technologies.

Some AI developers argue that the use of copyrighted content for training falls under fair use, especially if the training process is transformative and provides significant societal benefits. However, this interpretation is contested and varies by jurisdiction.

Read more Scarlett Johansson accuses OpenAI over chatbot voice

This controversy was brought to the fore at the end of 2023 when Ed Newton-Rex, the Head of Audio of AI developer Stability AI, resigned from his position due to disagreements over the company’s stance on the use of copyrighted material in training generative AI models. Newton-Rex, who has been a vocal advocate for the rights of creators, fundamentally disagreed with Stability AI’s interpretation of “fair use” in this context.

Stability AI, like many other AI companies, argued that their use of copyrighted materials is transformative and thus falls under fair use. This interpretation is based on the notion that the AI-generated outputs significantly alter the original works. However, Newton-Rex contended that this practice undermines the potential market for the original works and therefore does not meet fair use criteria. He emphasised that using copyrighted material without the creators’ consent or compensation is unjust and could harm the creative industries by replacing the demand for the original works.

In his resignation announcement, Newton-Rex called for better frameworks to protect creators’ rights, suggesting that AI companies need to obtain explicit consent from content creators before using their work for training models. He proposed various solutions, such as revenue-sharing agreements, upfront payments, or equity stakes, to ensure creators are compensated fairly. In 2024 he founded Fairly Trained, a not-for-profit company that provides certification for generative AI companies that operate according to fair training data practices.

Creator and technologist

Newton-Rex is a composer and entrepreneur who has been working in generative AI since 2010. His musical education started when he was a chorister in the Choir of King’s College, Cambridge, where he sang from the age of eight to 13. In 2010, having graduated from Cambridge University with a Double-Starred First in Music, he founded Jukedeck, the first startup in the field of AI music composition. He built the company’s first AI composition technology, with investment from Cambridge University. Jukedeck launched the first commercial AI music generation platform, which allowed non-musicians to easily create rights-cleared music in seconds. It was used to create more than a million pieces of music, that were used as backing in online videos, ads, TV and radio shows, podcasts and games. Jukedeck won several awards, including a Cannes Innovation Lion, and it was acquired by ByteDance in 2019.

At ByteDance, he led the AI Music Lab, then moved to Product Director for TikTok in Europe. After ByteDance, he led product for Voisey, a music creation app acquired by Snap. In 2022 he joined Stability AI, the generative AI company behind Stable Diffusion, to lead its audio team. They launched Stable Audio, Stability’s music generation platform, in September 2023, which was named one of TIME Magazine’s Best Inventions of the year.

Read more Navigating ethics and AI in broadcast production

Passion for AI

Newton-Rex’s interest in generative AI goes back more than a decade. “In 2010 I became obsessed with the question of whether it would be possible to teach a computer to compose music,” he says. “I thought that, if you could do it, there would be amazing applications of the technology in assisting and inspiring musicians, and in personalising music to the individual. I started Jukedeck because I thought it should be possible.”

One of the reasons Newton-Rex originally joined Stability AI was his desire to build a new music generation system, trained on licensed data. “I considered beginning a startup, but I read about Stability and thought we could get to market faster if I went and built it there,” he says. It was this strong belief in using licensed data for training that led to his split with the company.

“One of the four factors considered when deciding whether an act of copying falls under the fair use copyright exception is the [new work’s] effect on the potential market for or value of the copyrighted work,” he says. “This tends to get overlooked by proponents of generative AI training without rights holder consent. I think it is critical. Much generative AI competes with the work it is trained on. For me, this tips the balance away from fair use in many cases of generative AI training.”

Companies like Stability AI and OpenAI argue that without a broad interpretation of fair use, the development of their technologies will be hindered. Newton-Rex disagrees. “The exercise of copyright in this regard may slow generative AI development, but I believe you’ll ultimately reach the same endpoint with very capable models, just built in a sustainable way - not unfairly decimating the creative industries,” he says. “I think it’s worth taking a little longer to reach those highly-capable models to make the relationship between the AI industry and the creative industries symbiotic rather than predatory.”

Job loss

There are concerns in the industry of generative AI eliminating jobs. In the near term, producers of library music and background actors are already being replaced with computer-generated creations.

Read more MPTS: British drama and production jobs under threat

Newton-Rex takes a nuanced view of this trend. “In general, I want to stress that I don’t think it’s de facto bad for technology to replace work tasks,” he says. “What I object to in generative AI right now is people being replaced through exploitation of their work. Clearly there are other considerations, but I do think it’s important to make the distinction between generative that trains on people’s work without permission, and generative AI that takes a different approach. I fear that the current wave of AI will lead to huge job loss in the creative industries. I think this applies equally across all the creative industries. We are already seeing it in the fields in which generative AI is most advanced - particularly image generation. Artists are losing work and losing income. This will happen across all the creative industries. I have no doubt AI will be able to create meaningful, long-form content across all media types, and I honestly don’t think there are any limitations. This has been my view since I started working in this field.”

Fairly Trained

Fairly Trained, Newton-Rex’s new company provides Licensed Model (L) certification for AI providers. The L certification can be obtained for any generative AI model that doesn’t use any copyrighted work without a license. He describes his aims for the company: “If we’re successful, more generative AI companies will license their training data than do so today, and consumers and companies who want to use generative AI will have options that don’t exploit creators. I don’t think we’ll certify all the generative AI companies out there - but if we can encourage more AI companies to license their training data, I think we’ll be helping get a fairer deal for human creators.”

Where do we go from here?

“Generative AI is a very broad category, and as such I think it’s solving many different problems,” says Newton-Rex. “In my view, some are worth solving, and others aren’t. Generative AI is behind AI assistants, and these promise to be generally useful, potentially huge efficiency enhancers. But generative AI that’s used to flood the world with content, with very little human input, will I think do more harm than good. In general, I’m a fan of generative AI tools that only assist and don’t replace; and I’m opposed to generative AI tools that replace by letting people create content with a simple prompt.”

There is the old saw often used in terms of progress, ‘we can do it, but should we?’ Newton-Rex reflects on this question. “I’m really not sure,” he says. “For a long time, I thought we should and that’s why I worked on it. Now it’s harder to say. Again, I think that we should probably build some of these tools, but not all of them. We should build assistants, and models designed only to assist, and be thoughtful as we build them. But should we build content generators that will flood the market with middle-of-the-road music, text, images, videos and more? No, almost certainly not.”

Read more IBC2024 Primed to Lead Innovation, Explore Trends and Foster Collaboration across the Global Media Technology Community