AI Safety Researchers Call for Open Science and Greater Transparency

Summary

An open petition calls on developers of frontier AI models to share detailed safety methods and research with the global scientific community. It asks developers to provide enough information about each model’s safety-relevant properties for independent researchers to examine, challenge, reproduce, and improve the work. The requested disclosures include safety evaluations, safety-training recipes, relevant code and data, and evidence of both desired and undesired model behavior. The organizers argue that transparency would improve assessment by making the effectiveness and limitations of safety measures easier to study. They also say shared evidence could help universities and companies direct research funding toward the most promising safety approaches and spread useful practices across models, including systems that remain closed or open. The petition acknowledges that some information may create a credible security or misuse risk if released, but says exceptions should be specific and proportionate rather than a general justification for withholding safety evidence. The initiative was started by Martin Jaggi and Robert West of EPFL, Philip Torr of Oxford, and Anna Hedström of ETH Zurich. The page says it was published on September 18, 2026, will be updated over time, and lists endorsements from researchers and professionals who signed in a personal capacity.