Graceful degradation

A design property in which a system responds to a detected fault by moving to a reduced-functionality or safe operating state rather than failing catastrophically or continuing to operate on assumptions that may no longer hold. NASA's Mars rover fault-protection software, which places the vehicle into a conservative safe mode when it encounters an unresolved anomaly, is a documented example.

The design property of an AI system that responds to an input it cannot handle by flagging uncertainty and handing off to a human, rather than producing a confident guess. The intended answer to the brittleness question "what happens on an input it has never seen?"

A design pattern in which an agent, on detecting a reason to distrust its own next move, automatically narrows its own authority (for example, from submitting an action to only proposing it) rather than either continuing at full authority or stopping completely.

Designing an AI capability so a business process still works with reduced function when a dependency is unavailable (for example showing popular items when the recommendation service is down), rather than failing entirely.

A design in which a system steps down through pre-built, tested tiers of reduced capability when it cannot do its full job, rather than failing completely. The lowest tier ("the floor") is an honest fallback message.