In a world grappling with the pervasive nature of online hate speech, the potential of automated counterspeech has emerged as a beacon of hope. However, can AI truly combat hate effectively if it's constrained by safety measures? New research explores this crucial question, examining the impact of safety guardrails on the argumentative strength of large language models (LLMs) when generating counterspeech. The study delves into the delicate balance between "helpfulness" and "harmlessness" in LLMs, questioning whether these safeguards, designed to prevent harm, might inadvertently stifle the very qualities needed for persuasive counterspeech. Researchers used a dataset from a white supremacist forum to test how LLMs generated counterspeech with and without safety guardrails in place. They also experimented with different argumentative strategies, targeting specific components of hate speech like implicit stereotypes or focusing on the weakest points in the hateful arguments. The results reveal a surprising tension: safety guardrails, while intended to prevent the generation of harmful content, can actually weaken the cogency and persuasiveness of the counterspeech. LLMs without these constraints produced stronger arguments, directly addressing the hateful claims. This doesn't necessarily advocate for removing all safety measures, but it highlights the need for a more nuanced approach. The study emphasizes the importance of finding the right balance to ensure that automated counterspeech can be both safe and effective in combating online hate. The challenge lies in refining the way safety is implemented, allowing AI to push back strongly against hate speech while avoiding harm. This research paves the way for a more robust and effective use of AI in the fight against online hate, calling for a careful recalibration of the trade-off between safety and impact.