The fight against online hate speech is a constant battle, and researchers are exploring new ways to combat its spread. One promising avenue is the automatic generation of counter-narratives – reasoned responses that challenge hateful rhetoric. But how do you evaluate something as nuanced as a counter-narrative? New research tackles this very problem, proposing an innovative method using Large Language Models (LLMs) as judges. Traditional metrics like BLEU and ROUGE, commonly used to evaluate text generation, often miss the mark when it comes to the subtleties of counter-narratives. This research suggests using LLMs to compare counter-narratives head-to-head, creating a ranked tournament of different approaches. The results? This LLM-based ranking system correlates strongly with human judgment, showing promise for automating the evaluation of counter-narrative effectiveness. The research also tested various LLMs as zero-shot counter-narrative generators. Surprisingly, chat-aligned models like Zephyr outperformed instruction-tuned and base models, indicating their potential for directly combating online toxicity. While fine-tuning these models generally decreased performance, base models actually benefited from it. This suggests that refining base models with specific training data could be a valuable approach. However, challenges remain. The research highlights the importance of ensuring factual accuracy in counter-narratives, a critical area for future development. By refining these models and incorporating strategies like retrieval augmented generation, we may be able to create powerful AI tools to fight hate speech effectively.