Overcoming the Incentive Collapse Paradox

AI-assisted task delegation is increasingly common, yet human effort in such systems is costly and typically unobserved. Recent work by Bastani & Cachon (2025); Sambasivan et al. (2021) shows that accuracy-based payment schemes suffer from incentive collapse: as AI accuracy improves, sustaining positive human effort requires unbounded payments. We study this phenomenon in a budget-constrained principal-agent framework with strategic human agents whose output accuracy depends on unobserved effort. Our first contribution is a general impossibility result showing that incentive collapse is not merely a limitation of simple linear payments, but arises for any payment rule based only on observed task accuracy. To overcome this barrier, we propose a sentinel-auditing payment mechanism that enforces a strictly positive and controllable level of human effort at finite cost, independent of AI accuracy. Building on this incentive-robust foundation, we develop an incentive-aware active statistical inference framework that jointly optimizes (i) the auditing rate and (ii) active sampling and budget allocation across tasks of varying difficulty to minimize the final statistical loss under a single budget. Experiments demonstrate improved cost-error tradeoffs relative to standard active learning and auditing-only baselines.

AI systems are increasingly used to help people complete tasks such as data labeling, document review, and scientific analysis. However, when the AI is very accurate, human workers may have little incentive to carefully check its suggestions, because simply accepting the AI output will usually appear correct. This creates a problem: paying workers only based on observed accuracy can become extremely costly if we want them to continue putting in real effort. This paper studies how to avoid this problem. We propose a simple auditing method in which a small fraction of tasks are designed so that the AI assistant deliberately produces an incorrect output, allowing human effort to be directly rewarded. These special tasks allow the system to encourage careful human review without requiring payments to grow as AI systems become more accurate. We then develop a statistical framework that combines this auditing idea with task selection under a fixed budget. The framework helps organizations decide which data points should be labeled by humans, how often to audit, and how much to reward workers for passing audited tasks. Experiments on survey and biology datasets show that this approach can produce more reliable statistical conclusions at lower cost than standard methods.