Imagine a world where students in underserved communities can access personalized math tutoring right from their phones. That's the promise of AI-powered tutoring systems like Rori, a WhatsApp chatbot designed to help middle school students in countries like Sierra Leone and Liberia. But with the incredible potential of AI comes the responsibility to ensure these systems are safe and effective. A new study explores how researchers built and tested safety measures into Rori's growth mindset conversations, a key component of the tutoring experience. They found that while large language models (LLMs) like GPT-3.5 rarely generate harmful content on their own, the real challenge lies in moderating student input. Kids, being kids, might use inappropriate language or ask sensitive questions, and the AI needs to respond appropriately. This presents a unique challenge, blurring the lines between content moderation and classroom management in a digital space. The researchers tackled this by implementing a multi-layered safety system. First, a simple word filter catches obvious curse words. Then, a more sophisticated statistical model analyzes messages for various categories of harmful content, such as harassment or self-harm. Depending on the risk level, the system either prompts the student to rephrase or, in severe cases, ends the conversation and alerts a human moderator. But it's not just about catching the bad stuff. The study also highlighted the importance of designing conversations that encourage positive interactions. By structuring the growth mindset chats and keeping them focused, the researchers significantly reduced the risk of students straying into sensitive territory. This research underscores the need for a nuanced approach to AI safety in education. It's not enough to simply filter out toxic language; we must also teach AI tutors to respond thoughtfully and appropriately to the wide range of student behaviors, ensuring a supportive and productive learning environment.