npr.org
This is how evaluators test if an AI model is safe, and what they're checking for
Anthropic and OpenAI are expanding their work with third-party evaluators to make their products safer. But questions abound about the state of the science as well as the field's independence.