Large language models (LLMs) like ChatGPT sometimes 'hallucinate,' confidently making up facts. Think of it as the AI equivalent of sleep-deprived rambling. But how do you measure and fix these AI hallucinations at scale? New research introduces ANAH-v2, an innovative framework that tackles this challenge. ANAH-v2 is like an AI hallucination detective, carefully analyzing model outputs to identify and categorize factual errors. It works by building up a massive 'hallucination dataset' and training another AI model to spot inconsistencies. This process works in three stages: establishing a baseline, scaling the types of model errors it can identify, and finally, expanding the topics it covers. The result? A lean, 7-billion parameter model that outperforms even giants like GPT-4 in accurately detecting these factual flaws. The team found that AI models hallucinate more in Chinese than in English and generally perform better when given reference material (no surprise there!). Intriguingly, model size doesn't directly correlate with fewer hallucinations. DeepSeek's 67B model proved less hallucinatory without references, and Qwen1.5-14B excelled when using them. This work is a significant step toward more truthful AI. ANAH-v2 not only helps us measure the problem, it also offers solutions. By re-ranking model outputs based on 'hallucination scores,' it improves the accuracy of responses. This has profound implications for any field using LLMs, as it helps build trust and reliability. Future research will explore applying this to other tasks like dialogue generation and improving performance across different languages.