What Information Should Be Shared with Whom "Before and During Training"?

Pre-training transparency can improve risk management: The paper argues that sharing information about frontier AI models before training begins—such as expected compute usage, dataset descriptions, and capability assessments—allows for "shared awareness and preparedness" and "improved risk assessment and management" across stakeholders.

Different information requires different audiences: The paper distinguishes between information suitable for public disclosure versus information that should be shared only with "trusted actors" like home governments. This tiered approach aims to balance accountability with legitimate concerns about commercial sensitivity and security risks.

Key pre-training disclosures should include capability predictions and risk thresholds: Companies should communicate expected capability levels and assess whether their models will "approach any risk threshold," enabling governments and other actors to prepare for potential harms before training actually occurs.

Monitoring plans deserve advance scrutiny: The paper recommends that companies disclose their plans for "how the company will monitor progress, capabilities and risks during training," suggesting that oversight mechanisms themselves should be transparent and subject to review before training begins.

Pre-training commitments support accountability: Sharing specific details like training start/end dates and compute requirements creates a record against which companies' actual practices can be measured, contributing to "greater predictability and accountability" in frontier AI development.