OpenAI outlines AI safety checks

OpenAI has announced a set of rules for assessing the safety of its advanced AI models amid calls for greater external scrutiny from developers. The reviews will allow independent assessors to evaluate whether labs meet safety claims and provide more information to the public. The proposed reviews will cover four key areas: safety cases from training to deployment, defenses against misuse, and measurements of chemical and biological risks. These assessments will generally last weeks or months and will not be tied to individual product launches, although their findings could inform deployment decisions.