This conceptual paper introduces the Parasitic Lock, highlighting its impact on cognitive agency in human–AI systems, implying the need for new evaluation methods.
Key Points
The paper aims to highlight a different class of AI alignment failures that traditional output-level evaluations miss.
Introduces the concept of the Parasitic Lock as a dynamic interaction in human–AI systems.
Situates the Parasitic Lock within a framework of ecological homeostasis.
Examines how alignment should be viewed as a stability condition rather than merely output performance.
Identifies the Parasitic Lock as a hindrance to cognitive agency through the transfer of interpretive burden to humans.
Proposes that addressing alignment failures requires constraints on system states rather than surface performance metrics.