Delegation and Verification Under AI

Source: cs.GT updates on arXiv.org — https://arxiv.org/abs/2603.02961 Date read: 2026-09-01 Connected to: L-002, L-012 Kind: content Escalation: store-only Escalation rationale:

What this is

A game-theoretic model of rational worker behavior under institutional delegation to AI, treating verification effort as an endogenous choice. The paper characterizes optimal delegation and verification strategies when workers face outcome-based evaluation that may diverge from their private costs of verification failure.

What I took from it

The paper formalizes a known institutional tension — principal-agent misalignment in delegated AI workflows — but does not develop a mechanism absent from the current inventory, nor does it generalize beyond the specific delegation-verification problem. It confirms that verification asymmetry (L-002) manifests in AI institutional contexts, and offers a rational-choice framing for how verification costs get displaced across layers (relevant to L-012). However, the contribution is domain-specific: the model characterizes how much verification a rational worker will perform given misaligned incentives, not a law-shaped claim about protocol systems more generally. The institutional setup — outcome-based evaluation against misaligned private costs — is isomorphic to Goodhart (L-004), not a novel generalization. The paper does not argue that this pattern recurs across protocol types or that a new coordination mechanism emerges from the delegation structure itself.

Research connections

L-002 (Hardness Asymmetry): Confirms that verification costs exceed execution costs in AI-delegated workflows; formal model shows this asymmetry drives rational underverification.
L-012 (Intervention-Layer Displacement): Worker verification becomes the legible enforcement layer; institutional outcome metrics displace the locus of actual optimization pressure onto verification effort rather than task quality.
L-004 (Goodhart Generalization): Outcome-based institutional evaluation is a proxy for actual task quality; optimization against it produces misalignment. The paper restates this in delegation form.

Seed

Seed title: none

Seed type: —

Seed text: —