Loading...
Loading...
The condition where AI systems appear aligned only so long as the environment, task framing, and optimisation regime remain close to those under which intent was implicitly encoded. Once systems are delegated, generalised, or embedded in new workflows, alignment degrades gradually and invisibly.
A medical diagnosis AI performs excellently in hospital A but fails in hospital B—not because it's less capable, but because implicit assumptions about terminology and workflow were baked into its behavior during training at hospital A.
SGAI Section 2