OpenAI's New AI Safety Program Criticized for Falling Short of 'Employee-Like Access' Promise
OpenAI has announced a new third-party evaluator program aimed at enhancing AI safety, where staff from four AI safety and research organizations will serve as 'embedded assessors.' The company states these assessors will provide a clearer view of how AI labs evaluate and manage potential risks, with CEO Sam Altman emphasizing the value of outside opinions. However, many AI safety researchers are expressing significant concerns, arguing that the program falls short of Altman's September vow for independent evaluators to have 'employee-like access' to internal company systems. Critics, including Sydney Von Arx of Nightingale and former OpenAI employee Tomek Korbak, point to limitations such as narrow access scopes, restricted communication with employees, and no access to internal communication tools like Slack. OpenAI, in response, stated there are no current standards for embedded evaluations and expects to deepen these relationships over time.