Scaling, Lock-In, and Proxy Compliance: A Political Economy of Responsible AI

Technical Summary: Scaling, Lock-In, and Proxy Compliance

Problem Statement
The paper addresses a central puzzle in Responsible AI: the persistent coexistence of visible compliance (e.g., model cards, standardized evaluations, documentation) and inadequate substantive harm mitigation. While frameworks like the EU AI Act impose extensive documentation obligations, recurring harms and accountability gaps suggest that organizations may rationally choose "proxy compliance"—producing observable signals of responsibility while under-investing in actual risk reduction. The authors argue this is not merely a failure of ethics frameworks but a structural outcome of institutional parameters: specifically, the interplay between vendor-controlled evidence generation, deployer switching costs (lock-in), and enforcement mechanisms that rely on verifiable evidence.

Methodology
The authors develop a sequential political-economy game involving three agents: an AI vendor, a deployer, and a regulator. The model is static and single-vendor, treating scale and market concentration as the institutional background rather than endogenous variables.

• Timeline:

  1. The vendor chooses auditability ($t$) and substantive mitigation ($e$).
  2. The deployer observes these choices, decides on adoption, and if adopted, chooses a level of monitoring ($m$). Adoption incurs switching costs ($s$), representing lock-in.
  3. Harm ($h$) is realized. Evidence is generated with probability $p(t, m)$, which depends on both auditability and monitoring.
  4. Enforcement occurs if actionable evidence is found.

• Key Mechanisms:

  • Auditability ($t$): Helps satisfy procurement floors but also increases the probability of detection.
  • Mitigation ($e$): Directly reduces the probability of harm.
  • Lock-in ($s$): Reduces the deployer's effective recourse ($\delta_{eff}$), weakening the incentive to monitor.
  • Enforcement: Relies on verifiable evidence. The vendor faces penalties proportional to the probability of detection ($p + \Delta$), where $\Delta$ represents independent audit capacity.

• Equilibrium Concept: The authors solve for the subgame-perfect equilibrium where the vendor anticipates the deployer's monitoring response. The deployer monitors only if the expected benefit of recourse (reduced harm exposure) outweighs the cost of monitoring, a calculation heavily influenced by switching costs.

Key Contributions

1. Formalizing Proxy Compliance: The paper defines "proxy compliance" as an equilibrium where the vendor meets the minimum observable procurement floor ($t^* = \bar{t}$) but chooses mitigation ($e^*$) strictly below the social first-best ($e_{FB}$). This occurs when high switching costs dampen deployer monitoring, which in turn lowers the vendor's expected enforcement exposure, reducing the incentive for substantive mitigation.
2. Endogenous Monitoring: Unlike models where monitoring is exogenous, this framework allows the vendor to anticipate how their mitigation choices affect the deployer's monitoring intensity. Higher mitigation reduces residual harm, which can paradoxically reduce the deployer's incentive to monitor, creating a feedback loop.
3. Policy Levers Analysis: The authors derive comparative statics for four distinct institutional levers: independent audit rights, portability (reducing switching costs), incident reporting, and outcome-linked liability.

Results
The analysis yields a unique interior equilibrium and characterizes conditions for full mitigation (the corner solution where harm is eliminated).

• Proposition 4.4 (Existence of Proxy Compliance): In the absence of strong independent detection ($\Delta = 0$) and with sufficiently high switching costs ($s$), the equilibrium converges to proxy compliance. As $s \to \infty$, monitoring $m^* \to 0$, and mitigation $e^*$ drops to a level determined solely by the baseline auditability floor, often far below the social optimum.
• Proposition 4.5 (Accountability): Increasing independent detection ($\Delta$) or reducing switching costs ($s$) raises equilibrium mitigation. However, the effect of $\Delta$ is attenuated by the monitoring response (as mitigation rises, monitoring falls), though it never reverses the positive effect.
• Theorem 5.1 (Audit Rights): Sufficiently strong independent audit rights ($\Delta$) can eliminate proxy compliance and achieve full mitigation ($e^* = e_{FB}$) regardless of the level of lock-in ($s$). This is the most powerful single lever in the model.
• Proposition 5.2 (Portability): Reducing switching costs raises mitigation by restoring the deployer's leverage and incentive to monitor.
• Proposition 5.3 (Incident Reporting): Mandatory reporting creates an independent evidence channel ($\mu\rho$) that substitutes for vendor-gated access, raising mitigation even if auditability remains low.
• Proposition 5.4 (Outcome-Linked Liability): Liability tied directly to realized harm ($\lambda_{OFO}$) raises mitigation independently of the detection probability $p(t, m)$, providing a robust incentive when verification is weak.

Significance and Claims
The paper claims to provide a tractable institutional explanation for why "visible compliance and limited mitigation may coexist as rational choices." It argues that the gap between observable responsibility signals and harm reduction is not necessarily a failure of implementation but an equilibrium outcome driven by:

1. Vendor control over evidence: When vendors control the infrastructure for generating verifiable evidence (e.g., API access, logs), they can limit auditability.
2. Lock-in: High switching costs weaken the deployer's ability to act as a disciplining mechanism.
3. Evidence-dependent enforcement: When penalties depend on detected evidence rather than outcomes, vendors have an incentive to minimize transparency if the marginal cost of auditability exceeds the marginal increase in expected penalties.

The authors emphasize that their model is a benchmark, not a claim that any single governance instrument is sufficient. They caution that audit independence, reporting coverage, and credible sanctions are prerequisites for these mechanisms to work. The paper concludes that accountability requires both verifiability (independent access to evidence) and power to act (credible recourse for deployers), suggesting that procedural compliance (documentation) without these structural supports will likely result in persistent proxy compliance.