AI oversight is theoretical, not real — and the gap widens as labor gets displaced

This paper distinguishes between nominal oversight (humans officially in charge) and genuine oversight (humans actually able to understand and override AI decisions), arguing the distinction is absent from current AI governance rules. In practice, this means AI systems are treated as supervised when they're often opaque even to their supposed overseers — a gap that gets worse as AI replaces workers and concentrates decision-making power among fewer technical specialists.

Current AI governance assumes humans can meaningfully review and reject AI outputs. This paper argues that assumption is false for most deployed systems — the cognitive and technical barriers are too high. As AI handles more consequential decisions (hiring, lending, medical triage, resource allocation), those decisions concentrate among people with both technical understanding and institutional power to override the system, while everyone else gets formal oversight that doesn't actually work. This isn't a minor governance gap; it's describing a structural inversion where nominal accountability is theater while actual control narrows.