Most AI systems pursue goals assigned to them from the outside. Change the objective, adjust the reward signal, and the system simply optimises for something else; nothing about its own functioning is at stake. This paper asks whether a different kind of architecture is possible: one where the conditions for remaining in control are built into the system itself, rather than specified from outside. Affective Control under Uncertainty (ACU) formalises what this would require: a continuous viability-monitoring signal that globally shapes how the system selects actions, and a conditional self-modelling process that activates only when uncertainty, stakes, and time pressure converge. The paper specifies the architecture, proposes a discrimination experiment, and remains deliberately neutral on consciousness. The goal is to make a genuinely architectural question more precise and testable.
Scott McFarnell (Fri,) studied this question.