The generative AI boom is fueling a surge in demand for training data, but AI companies are facing growing challenges in obtaining it freely as data owners push back.
Key developments in the AI training data landscape: The music industry’s lawsuit against AI music companies Suno and Udio sends a strong message that high-quality training data is not free:
- Sony Music, Warner Music Group, and Universal Music Group are suing Suno and Udio for alleged copyright infringement, claiming the companies used copyrighted music in their training data “at an almost unimaginable scale.”
- The lawsuits could set a precedent for the future of AI music and the broader generative AI industry, as they highlight the growing leverage data owners have in the face of AI companies’ need for vast amounts of high-quality training data.
Shifting dynamics between AI companies and data owners: As the demand for training data soars and data becomes scarcer, AI companies are starting to strike deals with data owners:
- OpenAI has been making agreements with news publishers like Politico, the Atlantic, Time, and the Financial Times, exchanging money and citations for access to their news archives.
- YouTube announced plans to offer licensing deals to top record labels in exchange for music for training, signaling a shift towards a more collaborative approach between AI companies and data owners.
Potential implications for the AI industry: The changing landscape of training data access could have mixed consequences for AI development:
- AI companies may be forced to build smaller, more efficient models that are less polluting, or they may have to invest heavily in accessing data at the scale needed for larger models, potentially concentrating power in the hands of the biggest players.
- Introducing consent and compensation into the process of using training data could lead to a fairer data economy, giving data owners and individuals more agency in deciding how their data is used and allowing them to benefit from the AI boom.
Analyzing deeper: While the lawsuits and deals between AI companies and data owners are ushering in important changes, they also raise some concerns:
- News publishers may be making a Faustian bargain with AI, as language models are fundamentally incapable of being factual and often hallucinate citations, making it difficult for AI companies to honor their promises of attribution.
- The concentration of power in the hands of the biggest AI players, who can afford to pay for vast amounts of training data, could stifle competition and innovation in the industry.
Ultimately, the evolving landscape of AI training data highlights the need for a more equitable and consensual approach to data use, one that balances the interests of AI companies, data owners, and individuals while fostering responsible innovation.
AI companies are finally being forced to cough up for training data
