The safe harbor paradox: when combining human and AI skills creates or destroys value

Human oversight is widely assumed to mitigate the risks of deploying artificial intelligence in complex tasks. This paper challenges that assumption, showing that the human-machine skill policy is a high-variance proposition: it yields the highest economic utility when genuine augmentation is achieved, yet destroys value when it is not. We call this the safe harbor paradox. The paradox is robust: it persists in high-error-cost environments even when machine skills uniformly outperform human skills, and is not resolved by improving cost-effectiveness over time. It also survives stochastic perturbation of the economic chain downstream of performance, at levels well beyond plausible operating variability. Augmentation is the joint achievement of two channels: routing work between the two skills at each execution, and the synergy a workflow adds. The stake of routing is set by the skills alone and splits at the better of the two: a task-level part, secured by an ex ante commitment and only as good as the assessment behind it, and a residual advantage, reachable only by weighting the skills instance by instance and identifiable only from paired observation of both. When augmentation fails, the loss has two separable drivers: the dual-cost margin structure is dominant, yet the error regime alone sustains the paradox where errors are expensive and the two skills are comparable. Through Monte Carlo simulation, we map the conditions under which each skill policy succeeds or fails, giving a tool for ex-ante evaluation and screening. Our findings carry implications for policymakers, as blanket oversight mandates may create false security; for organizational leaders, as augmentation demands deliberate design; and for workers, as substantive human roles must be distinguished from performative ones.