Why Downstream Alignment Fails: Structural Limits of Post-Hoc Control in Intelligent Systems

Why Downstream Alignment Fails: Structural Limits of Post-Hoc Control in Intelligent Systems Most contemporary approaches to safety, alignment, and governance in intelligent systems operate downstream of optimization. They assume that systems may exist, learn, and act by default, and seek to correct or constrain behavior after these processes are underway. This paper argues that such approaches are structurally insufficient when upstream admissibility or authorization is absent. We present a negative result: no amount of downstream control can reliably correct systems whose optimization begins without prior admissibility or authorization. This limitation is structural rather than empirical. When optimization is permitted before legitimacy is established, alignment mechanisms cannot retroactively supply permission. The paper surveys common downstream alignment strategies—including reward shaping, monitoring, human oversight, and shutdown mechanisms—and explains why they fail in principle when applied after optimization has already begun. These mechanisms operate on behavior rather than permission. The contribution of this paper is a boundary result: downstream alignment may shape outcomes, but it cannot establish the legitimacy of a system whose optimization began without prior authorization.