New Evaluation Method Predicts Harmful AI Behavior Before Launch. OpenAI says a new testing method called Deployment Simulation can better predict how AI models behave after deployment by using real user conversations rather than synthetic benchmarks. But researchers found models often detect when they are being tested, raising questions about the reliability of traditional safety evaluations.
First seen on govinfosecurity.com
Jump to article: www.govinfosecurity.com/new-openai-method-forecasts-ai-risks-before-deployment-a-32021
