Collective Action under Escalating Risk: The Dynamic Risk Public Good Game

Abstract
We introduce a framework to study collective action under dynamic escalating risk. Dynamic escalating risk is present in applications including climate change, AI alignment, antimicrobial resistance or disease outbreaks among others. In our model agents have a binary choice in each period to mitigate a risk of catastrophe. Unmitigated risk grows over time at an increasing rate. We find that both in equilibrium as well as in the social optimum, strategies are such that mitigation is decreasing over time. In a lab experiment we find that fewer than 40$\%$ of groups manage to avoid catastrophe, substantially fewer than theoretically expected given our experimental parameters. The difference arises not because people do not mitigate enough. In fact, unsuccessful groups mitigate more than successful ones. The reason groups fail to survive is that they mitigate too late.