The European Union’s AI Act is set to establish a unified regulatory framework for artificial intelligence across the EU, introducing significant new requirements specifically for generative AI. Notably, it mandates the disclosure of copyrighted works used in training these systems, raising important implications for AI developers and providers.
Background on the AI Act Amendments
The recent amendments to the AI Act specifically target generative AI, requiring providers to publicly disclose details about copyrighted works used during the training of their models. This legislative move addresses growing concerns over copyright infringement in the training processes of AI systems, which are capable of generating text, images, and other media. The rapid growth and investment in AI technologies, highlighted by OpenAI’s ChatGPT and Microsoft’s substantial AI investments, underscore the urgency of these regulations.
Many AI systems rely on vast amounts of copyrighted material, often scraped from the internet, leading to allegations of infringement from rights holders. The new disclosure requirements are expected to increase the legal scrutiny faced by AI developers, particularly within the EU.
New Obligations for Generative AI Providers
Compliance Policies
Providers of general-purpose models must establish policies to comply with EU copyright laws. This includes respecting opt-outs from the EU’s commercial text and data mining (TDM) exception using state-of-the-art technologies to avoid infringement.
Global Scope of Compliance
The AI Act requires adherence to EU copyright laws, even if models are trained outside the EU. This ensures a level playing field and prevents providers from gaining a competitive advantage by operating under lower copyright standards in non-EU countries.
Disclosure Requirements
Providers must publicly disclose details about the content used for training their models. This includes listing the main data collections or datasets, such as large private or public databases, while also protecting trade secrets and confidential business information. Notably, no exceptions are made for open-source models, and those fine-tuning general-purpose models must disclose new training data sources.
Broader IP Considerations
The AI Act aims to balance innovation with the protection of fundamental rights, including IP rights. While the legislation emphasizes protecting the intellectual property of AI developers, such as trade secrets, it also mandates public disclosure under specific conditions. Suppliers must provide necessary information for compliance without compromising their IP or trade secrets.
Impact on the UK and Future Developments
Following Brexit, the UK’s approach to AI regulation has diverged from the EU’s framework. The UK has not implemented the EU’s text and data mining exceptions, limiting such exceptions to non-commercial research. While the UK Government is consulting on broader exceptions, it has not yet proposed obligations for AI providers to disclose copyrighted works used in training.
The UK aims for a lighter regulatory touch to foster innovation in the AI sector. However, this approach may face challenges in a globalized market where the EU’s stringent regulations could set a de facto standard, impacting the UK’s competitiveness in AI development.
Actionable Steps for UK AI Developers
UK-based AI developers should collaborate closely with IP lawyers to navigate the complex landscape of training data and copyright compliance. Key steps include:
- Identifying necessary works and applicable copyright restrictions.
- Evaluating the extraction of protected data and potential infringement risks.
- Considering jurisdiction-specific strategies for training and hosting models.
- Implementing safeguards, such as filters and human oversight, to prevent infringing outputs.
- Meticulously documenting compliance efforts to adhere to regulatory requirements.
The EU AI Act represents a significant advancement in regulating generative AI and protecting intellectual property rights. AI providers must adapt to these new requirements by carefully considering where and how to train their models to ensure compliance with EU laws. This will involve detailed documentation, transparent disclosure of training data, and robust compliance policies to navigate the intricate landscape of AI and IP. The UK’s differing regulatory approach further emphasizes the need for thorough legal guidance and strategic planning in AI development and deployment.
