OpenAI's New AI Safety Program Faces Skepticism Over Limited Evaluator Access
OpenAI has announced a new AI safety program that involves staff from four AI safety and research organizations becoming 'embedded assessors' to evaluate potential risks. CEO Sam Altman stated that getting outside opinions on safety work and system robustness is increasingly important. However, many AI safety researchers are expressing concerns that this new program falls short of Altman's previous promise of 'employee-like access' for independent evaluators. Critics, including former OpenAI employees, argue that the assessors have very narrow scopes of access, limited interaction with employees, and no access to internal communication tools like Slack. This limited access is seen as a significant drawback, potentially hindering the evaluators' ability to identify blind spots and comprehensively assess the company's AI safety measures. OpenAI, in response, stated that there are currently no standards for embedded evaluations and that they expect to deepen these relationships over time.