Imagine AI grading essays, not on grammar or facts, but on *equity*. That's the fascinating premise explored by researchers at Carnegie Mellon University. They've developed an online lesson to train tutors on how to handle potentially inequitable situations students face, like lacking internet access at home. But grading nuanced responses about fairness and social justice is tough. So, they turned to the power of GPT-4. This study explored whether AI could accurately assess tutor responses to complex scenarios involving students like Jeremiah, who couldn't complete his homework due to lack of internet, or Alexis, struggling to hear the teacher from the back of the classroom. The AI was tasked with determining if tutors were effectively helping students advocate for themselves. While the AI generally performed well, especially when given a few examples beforehand (few-shot learning), it occasionally stumbled over subtle interpretations. For example, the AI considered the phrase, "It’ll help Jeremiah learn to take agency over his life," as demonstrating tutor understanding of student advocacy, while human graders disagreed. This highlights the challenge of teaching machines the nuances of human empathy and fairness. The researchers also compared the cost and speed of different GPT models. Interestingly, the slightly older GPT-4o, with a little help from few-shot learning, emerged as the sweet spot for balancing performance, cost, and efficiency. This research opens exciting doors for using AI in large-scale tutor training. Imagine personalized feedback for thousands of tutors, helping them support students facing diverse challenges – all powered by AI. This also raises important questions about how we define and assess "equity" in an increasingly AI-driven world. The data from this study is publicly available, inviting further exploration into the fascinating intersection of AI, education, and social justice.