In light of recent legal disputes where AI companies have faced accusations of using copyrighted content without consent for training their AI models, a novel certification has been introduced to identify copyright-compliant AI.
The Challenge
Legal challenges have arisen as AI companies employ various methods to train their models, including scraping data without permissions, as exemplified by the New York Times’ lawsuit against OpenAI. This lawsuit, claiming unauthorized use of New York Times articles, raises concerns about how AI companies source their training data.
The Initiative by ‘Fairly Trained’
‘Fairly Trained,’ a nonprofit initiative, has launched a Licensed Model (L) certification for AI providers. This certification is awarded to generative AI models that refrain from utilizing copyrighted material without a proper license. The initiative aims to emphasize the importance of securing rights-holder consent for generative AI training.
Who Qualifies?
The certification is open to any organization, entity, or product offering generative AI models or services, provided they meet specific criteria.
Certification Criteria
Key criteria for obtaining the certification include:
- Ensuring that training data is either explicitly provided to the model developer for training purposes, available under an appropriate open license, in the public domain, or fully owned by the model developer.
- Implementing a rigorous due diligence process to assess the rights position of the training data provider.
- Maintaining comprehensive records of the training data used for each model.
Certification Costs
In addition to meeting the criteria, AI companies are required to pay for the certification. The fee structure is based on the organization’s annual revenue, ranging from a $150 submission fee and a $500 annual certification fee for entities with a $100k annual revenue to a $500 submission fee and a $6,000 annual certification fee for organizations with a $10M annual revenue.
Adapting to Changing Practices
If a certified organization alters its data practices after certification, it is the organization’s responsibility to notify ‘Fairly Trained’ of the change. While no proactive monitoring is in place, ‘Fairly Trained’ reserves the right to withdraw certification without reimbursement if new information surfaces indicating that an organization no longer meets the criteria.
Challenges for Text-Based Models
A notable challenge with the certification is its applicability to text generation models. ‘Fairly Trained’ acknowledges that most major text generation models would likely struggle to meet the criteria due to their extensive use of copyrighted work. This dilemma arises from the broad scope of copyright protection, encompassing various content types such as blog posts, images, code, government documents, and more.
Exploring Alternatives
Ed Newton-Rex, CEO of ‘Fairly Trained,’ expresses hope that future models may require less copyrighted material for training. Potential alternatives include using open datasets, public domain works, generating synthetic data, crowdsourcing, utilizing data with appropriate licenses like Creative Commons, collaborating with creators, or implementing web scraping with strict filters.
Collaboration and Agreements
AI companies could also collaborate with artists, musicians, and creators to generate original content for training. Entering agreements with individuals or organizations for the use of private or proprietary data under suitable terms is another viable option.
Key Takeaways
Navigating the complexities of copyright compliance in AI training remains a nuanced issue. While the ‘Fairly Trained’ certification offers a commendable solution to differentiate ethical training practices, its applicability to text-based models poses a challenge. Furthermore, relying on organizations to self-report changes in their training data practices may require additional monitoring to ensure effectiveness.
Ultimately, ongoing legal developments and evolving industry practices will shape how AI companies approach their training methodologies in the future