The Competence Paradox

Assistance from AI can raise the quality and speed of a team's output while the people producing it lose the ability to notice when the system is wrong. This paper argues that the right target is appropriate reliance, meaning accepting good advice and rejecting bad, and that this capability rests on the very expertise routine reliance erodes. It separates trust, which is an attitude, from reliance, which is a behaviour, from decision quality, which is the outcome, and shows that adoption programmes measure the first two because low usage is visible while undetected error is not. Drawing on four decades of automation research, it sets out a five-stage erosion loop in which good output reduces checking, reduced checking reduces practice, practice loss erodes skill, and a plausible wrong answer then meets a weakened check. Every stage except the last resembles success. Four retention practices and four board measures follow, covering override rate, error-catch rate, unaided performance and tool-off drills. The practices are proposals rather than validated instruments, and the thresholds must come from each organisation's own baseline.