AI Safety Researchers Call for Open Science: Making Independent Review the Norm

Overview

Recently, dozens of researchers from prestigious academic institutions including ETH Zurich, EPFL, and the University of Oxford launched a petition titled “AI Safety Research Calls for Open Science.” This initiative strongly urges developers of frontier artificial intelligence models to share their AI safety methods and research findings with their global community, aiming to advance the field through independent scientific review.

“As AI systems become increasingly powerful, their safety should be documented in detail, including the methods and evaluations that enable independent researchers to examine, challenge, reproduce, and improve them.”

Launched on September 18, 2026, the initiative was spearheaded by scholars such as Martin Jaggi, Robert West, Philip Torr, and Anna Hedström, emphasizing that transparency will have a profound impact on assessing and enhancing AI model safety.

The Necessity of Open Science

The petition points out that the training data, safety fine-tuning recipes, and related code for many frontier models are currently not made public, hindering the scientific community’s deep understanding of model behavior. By sharing sufficiently detailed safety-related information, including evaluation results, safety training strategies, and evidence of desired and undesired behaviors, it is possible to significantly enhance awareness of the effectiveness and limitations of existing safety measures.

Furthermore, transparency helps channel research funding and human resources in academia and industry worldwide toward the most promising research directions. By leveraging the expertise and positive contributions of the global scientific community, sharing safety best practices can improve the safety of all AI models, regardless of whether they are open-source or originate from which country. This collaborative model is expected to accelerate progress in AI safety research, enabling it to keep pace with the rapid development of AI capabilities.

Exceptions and Supporting Signatories

While advocating for openness, the petition acknowledges exceptions in certain circumstances. If releasing certain information itself could constitute a credible security or misuse risk, exemptions may be granted. However, such exceptions should be specific and proportionate, rather than serving as a blanket excuse to keep safety evidence confidential indefinitely.

The initiative has garnered signatures from numerous researchers at top global universities including MIT, Harvard University, Princeton University, Imperial College London, and the National University of Singapore. The sponsors emphasize that all endorsements are made in an individual capacity, and listed affiliations are for identification purposes only. As more experts join, this call for the spirit of open science is gaining momentum, aiming to build a more transparent and trustworthy AI development environment.