Researchers Call on AI Labs to Disclose Safety Methods

🛡 Researchers Call on AI Labs to Disclose Safety Methods

On September 18, 2026, a group of scientists from EPFL, Oxford, and ETH Zurich published the letter “A Call for Open Science in AI Safety.” They call on developers of frontier AI models to publish evaluations, safety-training recipes, code, and data — so that independent researchers can verify them.

🌍 The focus is shifting from secrecy to verifiability: publish methods and evidence, not model weights. Risk-sensitive information can be withheld, but exceptions should be targeted, not a blanket excuse.

👤 “AI Safety” means specific documents: lists of evals, training recipes, code, and data. The letter was signed by approximately 208 researchers, including staff from Google DeepMind, Hugging Face, and EleutherAI; the document is promised to be updated.

Source 1: https://make-safety-open.github.io/ Source 2: https://news.ycombinator.com/item?id=49796651