A Call for Open Science in AI Safety

Sign the petition

We call on frontier model developers to share their AI safety methods and research with the global community, to enable independent scientific scrutiny and advancement.

The safety of increasingly powerful AI systems should be documented with detailed descriptions of methods and evaluations that independent researchers can examine, challenge, reproduce and improve.

Transparency will make a big difference for both assessing and improving the safety of AI models.

Assessing AI safety will be facilitated by sharing sufficient detail about the safety-relevant aspects of each frontier model—including evaluations, safety-training recipes, relevant code and data, and evidence of desired and undesired behavior. This will help advance the scientific understanding of model behavior and the effectiveness and limitations of safety measures.

Improving AI safety will be facilitated as transparency will allow research efforts and funds—across academia and industry—to be spent towards the most promising directions in a targeted way. Drawing on the expertise and active contribution of the global scientific community, sharing safety best-practices helps improve the safety of all AI models, whether closed or open, irrespective of geographic origin. This can accelerate progress in AI safety, allowing safety research to keep pace with the fast advances in AI capabilities.

In cases where releasing particular information could itself create a credible security or misuse risk, exceptions can be made. Such exceptions should be specific and proportionate, rather than becoming a general reason for keeping safety evidence closed.

Sign the petition

Signatories

Initiated by: Martin Jaggi (EPFL), Robert West (EPFL), Philip Torr (Oxford), and Anna Hedström (ETH Zurich).
Published: 18th Sept 2026
Will be updated from time to time. All endorsements are in a personal capacity. Affiliations are provided for identification only.