The Need for Frontier AI Transparency

As AI technology advances rapidly, securing transparency to ensure public safety and corporate accountability has become urgent. Until full safety standards are established, we propose a Targeted Transparency Framework as an intermediate step toward safely and responsibly developing the most powerful AI models.

This framework aims for flexible, lightweight regulation so as not to stifle innovation, and applies only to the largest-scale model developers rather than all developers. The applicable criteria consider computing power, cost, evaluation performance, annual revenue, and R&D spending, and include appropriate exemptions to reduce the burden on startups.

The core components are as follows:

• Establishing a Secure Development Framework: setting up procedures to assess and mitigate chemical, biological, radiological, and nuclear (CBRN) risks and model autonomy misalignment risks.
• Publishing the framework: disclosing the Secure Development Framework so that researchers and governments can verify the safety of the models.
• Issuing a System Card: releasing, at the time of deployment, a document summarizing testing and evaluation procedures, results, and mitigation measures.
• Whistleblower protection: strengthening substantive legal accountability by classifying the spread of false statements about compliance with the framework as a legal violation.