ResearchPod Summary
Responsible AI governance often relies on checklists, model cards, and standardized evaluations. Despite these efforts, post-deployment harms persist. This paper investigates why visible compliance signals frequently coexist with low levels of substantive harm mitigation. The authors argue that this is not merely a failure of implementation, but a rational equilibrium outcome in a market where vendors control the evidence-generating infrastructure and deployers face significant switching costs.
The authors develop a sequential game involving three primary actors: a vendor, a deployer, and a regulator. The vendor chooses a level of auditability (to meet procurement floors) and substantive mitigation (to reduce harm). The deployer, after integrating the AI system, decides how much to monitor the system for potential issues.
Crucially, the model incorporates "lock-in"—the idea that once a deployer has integrated a model into their workflows, switching to a different provider is costly. This lock-in reduces the deployer's incentive to monitor, as they have less credible recourse if they find evidence of harm. Because enforcement depends on verifiable evidence, the vendor anticipates this reduced monitoring and strategically limits their investment in mitigation, settling for "proxy compliance"—doing just enough to satisfy visible requirements without addressing underlying risks.
The study evaluates several policy interventions to break this cycle:
Ultimately, the authors demonstrate that while documentation and standards are necessary, they are insufficient for accountability if they do not address the underlying power asymmetries between vendors and deployers.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.