Unveiling the Truth: AI Firm's Perspective on Copyright Transparency

Unveiling the Truth: AI Firm's Perspective on Copyright Transparency

Table of Contents:

1. Introduction
2. The AI Act: Overview and Objectives
3. Foundation Models: Definition and Significance
4. The Current Obligation on Providers of Foundation Models
5. Argument 1: The Extensiveness of Copyright Protection
   • Case law examples
   • Challenges in documenting and disclosing copyright-protected data
6. Argument 2: Subjectivity of Copyright Protection
   • Lack of a copyright register
   • Differences in court rulings
7. Proposed Changes to the Obligation
8. Implementing Copyright Compliance in AI Projects
   • ML6's approach to acquiring licenses and assessing lawful access
   • Benefits of a comprehensive risk assessment for AI systems
9. Looking beyond the AI Act: Ethics and Trustworthiness
10. Conclusion

The Challenge of Copyright Compliance in the AI Act

The AI Act, currently undergoing final discussions, aims to regulate the risks associated with AI systems and impose obligations on providers of foundation models. Foundation models play a crucial role as large-Scale AI models trained on vast amounts of data, serving as the basis for further specialization and application in various domains. One of the proposed obligations is for providers to document and publicly disclose detailed summaries of training data protected under copyright law. While transparency is essential, implementing this obligation presents significant challenges.

Argument 1: The Extensiveness of Copyright Protection

Copyright protection encompasses a wide range of content, from books to images, text snippets, and even design objects. European court rulings have shown that even short phrases of 11 words can be protected by copyright. Considering that foundation models are trained on vast datasets, it becomes practically difficult to determine where the obligation to document and disclose copyrighted data should start and end. Implementing this obligation would Create enormous administrative burdens for providers.

Argument 2: Subjectivity of Copyright Protection

Determining whether content qualifies for copyright protection is subjective and relies on judicial interpretation. With no copyright register in place, judges are responsible for ruling on copyright protection in litigation cases. This subjectivity makes it challenging for providers of foundation models to assess which data is copyright-protected. Under the current version of the AI Act, providers would be required to make judgments on copyright compliance, a responsibility better suited for the judiciary.

Proposed Changes to the Obligation

Recent updates on trialogue discussions suggest that the obligation regarding copyright is under review. The proposed mechanism aims to focus on demonstrating that measures to respect law and ethics are in place, rather than identifying and disclosing specific copyright-protected works. This approach would provide a more practical and sensible framework for compliance, relieving providers of foundation models from the burden of extensively documenting copyrighted data.

Implementing Copyright Compliance in AI Projects

Ml6, an AI service provider, has already been proactively implementing copyright compliance measures in their projects. They prioritize acquiring licenses for using copyrighted works and rely on exceptions like text and data mining when licenses are unavailable. This approach emphasizes the necessity of lawful access and rights availability. Furthermore, ml6 recognizes the broader importance of addressing risks such as bias, discrimination, and data privacy in AI systems. A comprehensive risk assessment helps in identifying these risks and implementing appropriate measures.

Looking beyond the AI Act: Ethics and Trustworthiness

While the AI Act addresses important aspects of AI regulation, entities involved in AI development should proactively incorporate ethics and trustworthiness principles. Even before the implementation of the AI Act, organizations can start thinking about compliance measures and risk assessments in areas such as copyright, bias, and data privacy. Implementing robust processes and documentations enhances transparency and builds trust in AI systems.

In conclusion, the current obligation in the AI Act regarding copyright compliance poses significant challenges for providers of foundation models. However, recent discussions suggest potential changes that Align with a more practical approach, focusing on demonstrating compliance measures rather than exhaustive documentation. Implementing comprehensive risk assessments and ethical practices further contribute to the trustworthy development and deployment of AI systems.