Quality Assessment of Public Summary of Training Content for GPAI models required by AI Act Article 53(1)(d)

A team of researchers has developed the first formal framework for evaluating whether AI companies are properly disclosing what's in their training data, as required by the EU's landmark AI Act. The paper, authored by Dick A. H. Blankvoort, Harshvardhan J. Pandit, and Maximilian Gahntz, addresses Article 53(1)(d) of the regulation, which mandates that providers of general-purpose AI (GPAI) models like GPT-4, Claude, and Llama publish detailed summaries of their training content. The goal of this legal requirement is to give copyright holders, data protection authorities, and the public a clearer view into the datasets powering these influential systems.

The researchers' quality assessment framework examines public summaries across two critical dimensions: transparency (whether information is clear, comprehensive, and detailed) and usefulness (whether stakeholders can actually use the document to exercise rights related to IP, copyright, and data protection). This structured approach allows for comparing practices across different AI providers and helps regulators like the EU's AI Office identify systemic issues. The team applied their framework to the 5 public summaries available as of January 12, 2026, though the paper doesn't name the specific models or companies assessed.

To make their findings actionable, the researchers are developing a public website to share assessments, methodologies, and recommendations. This resource will serve both authorities enforcing the AI Act and AI providers seeking to create higher-quality disclosures. The work represents a crucial step toward operationalizing the AI Act's transparency provisions, moving from vague requirements to measurable standards that could shape how companies like OpenAI, Anthropic, and Meta document their training processes for regulators and the public.