'Fairly Trained' aims to incentivize ethical data sourcing practices by highlighting companies that legally license their training data. The certification process relies on written submissions detailing the dataset contents, their origin, and the company's due diligence processes. However, this trust-based system has been criticized for its potential loopholes, as companies often do not disclose detailed information about their datasets. Despite this, Newton-Rex believes that the certification will encourage better practices in the AI industry.
Key takeaways:
- Ed Newton-Rex resigned from Stability AI over disagreements about the use of copyrighted work to train AI systems. He launched a nonprofit, Fairly Trained, which certifies AI companies that only use data from consenting creators.
- Several lawsuits have been filed against AI companies like OpenAI, Meta, and Stability AI for allegedly using copyrighted work without permission to train their AI models.
- AI companies are arguing that using data scraped from the internet is "fair use" and doesn't violate copyright laws. However, some creators are fighting back by using tools that can "poison" AI models trained on their work without consent.
- Most large AI companies do not disclose the contents or details about the datasets used to train their models, making it difficult for creators to know if their work has been used without their consent. OpenAI has licensed news articles from several newsrooms for use as training data and is in discussions with others.