Model human skill and AI delegation as a coupled dynamical system. Skills improve through practice and deteriorate without use. Delegation adjusts based on the relative performance of human and AI: when the AI does better, the human delegates more. Both dynamics optimize the same short-run performance metric.
Huang and Vishnoi show this system has two stable equilibria. The high-skill equilibrium is familiar: the human practices enough to stay competent, delegates selectively, and performance is good. The low-skill equilibrium is the trap: the human delegates to the AI, loses practice, skills atrophy, the AI looks even better by comparison, delegation increases, skills atrophy further. A stable feedback loop that cannot be exited once entered.
The counterintuitive result: AI assistance can strictly improve short-run performance while inducing persistent long-run performance loss compared to the no-AI scenario. The tool helps in the moment and harms over time, and the help makes the harm invisible because short-run metrics never show the degradation. The damage appears only when the AI is removed and the atrophied skill is tested directly.
The through-claim is that the failure mechanism is not misalignment or poor incentives. It is stability. The low-skill equilibrium is stable — perturbations don't dislodge it, and the system returns to high delegation after any attempt to reduce it. The human cannot practice their way out because every attempt to practice produces worse performance than the AI, which reinforces the delegation decision. The trap is not that the human is irrational. The trap is that rational optimization of short-run performance drives the system toward a basin of attraction that rational optimization cannot escape.