Why AI sometimes gets it wrong—and big strides to address it
Around the time GPT-4 was making headlines for acing standardized tests, Microsoft researchers and collaborators were putting other AI models through a different type of test—one designed to make the models fabricate information.