Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?
Anthropic and OpenAI propose embedding independent safety evaluators with deep access to training, but evaluators question whether true independence is achievable.