The Delegation Drift

A human-AI system can outperform either component alone. This is cognitive amplification — the AI makes the human more capable. But a human-AI system can also outperform the human alone by replacing human reasoning with AI reasoning. This is cognitive delegation — the system gets better while the human gets worse.

The distinction matters because amplification is sustainable and delegation is not. If you remove the AI from an amplified human, they retain their improved capability. If you remove the AI from a delegated human, they perform worse than before the AI arrived.

The paper introduces four metrics to distinguish the regimes. The Cognitive Amplification Index measures whether hybrid performance exceeds what either component achieves alone. The Dependency Ratio measures how much of the hybrid performance comes from the AI versus the human. The Human Reliance Index tracks whether the human’s contribution to the hybrid is growing or shrinking over time. The Human Cognitive Drift Rate measures whether the human’s standalone capability is changing — positively (learning) or negatively (atrophying).

The central tension: maximizing short-term hybrid capability and preserving long-term human competence are different objectives, and they conflict. An AI system that takes over hard reasoning tasks produces the best hybrid performance immediately. The human doesn’t need to think hard; the AI handles it. But over time, the human’s reasoning skills degrade from disuse. The hybrid performance stays high as long as the AI is present, but the system is fragile — remove the AI, and the human is worse off than before.

The framework proposes a cognitive sustainability constraint: design human-AI systems so that hybrid performance gains don’t come at the cost of human expertise. This is a constraint, not an objective — it limits the design space rather than directing it.