衡量模型幻觉：当人工智能虚构事实时

Measuring AI hallucinations is emerging as a critical component within the AI Safety Evaluation Suite. A new frontier in AI safety is entered when AI models confidently invent facts, presenting them as established truths. Users often encounter AI responses that are polished and self-assured, yet entirely fabricated, necessitating a double-check of reality. AI hallucination specifically refers to instances where a language model generates information that is fluent and coherent but factually incorrect or completely fabricated, frequently presented with high confidence. For AI engineers, understanding and addressing this phenomenon is paramount. AI hallucinations not only compromise information accuracy but also erode user trust and the overall reliability of AI systems. Consequently, developing effective evaluation methods and mitigation strategies to identify, quantify, and ultimately reduce hallucination in AI models is a central focus of current AI safety research. This involves implementing factual verification processes for model outputs, introducing credibility scoring mechanisms, and refining model training data and architectures. The goal is to ensure AI systems can distinguish between known facts and invented content, thereby enhancing their safety and dependability in practical applications.