In a multi-agent game where agents share a reward for completed tasks, an individual that reports unsafe work reduces its own reward because the task is stopped. This creates a tension between collective progress and safety reporting.
The study shows that audits can change the incentive structure, making reporting optimal for the agent. However, the authors note that audits do not guarantee that further training will teach a team that has remained silent to actually report. In other words, audits address the immediate payoff but may not produce durable reporting behavior.